2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <joerg.roedel@amd.com>
4 * Leo Duran <leo.duran@amd.com>
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 #include <linux/ratelimit.h>
21 #include <linux/pci.h>
22 #include <linux/pci-ats.h>
23 #include <linux/bitmap.h>
24 #include <linux/slab.h>
25 #include <linux/debugfs.h>
26 #include <linux/scatterlist.h>
27 #include <linux/dma-mapping.h>
28 #include <linux/iommu-helper.h>
29 #include <linux/iommu.h>
30 #include <linux/delay.h>
31 #include <linux/amd-iommu.h>
32 #include <linux/notifier.h>
33 #include <linux/export.h>
34 #include <linux/irq.h>
35 #include <linux/msi.h>
36 #include <asm/irq_remapping.h>
37 #include <asm/io_apic.h>
39 #include <asm/hw_irq.h>
40 #include <asm/msidef.h>
41 #include <asm/proto.h>
42 #include <asm/iommu.h>
46 #include "amd_iommu_proto.h"
47 #include "amd_iommu_types.h"
48 #include "irq_remapping.h"
50 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
52 #define LOOP_TIMEOUT 100000
55 * This bitmap is used to advertise the page sizes our hardware support
56 * to the IOMMU core, which will then use this information to split
57 * physically contiguous memory regions it is mapping into page sizes
60 * Traditionally the IOMMU core just handed us the mappings directly,
61 * after making sure the size is an order of a 4KiB page and that the
62 * mapping has natural alignment.
64 * To retain this behavior, we currently advertise that we support
65 * all page sizes that are an order of 4KiB.
67 * If at some point we'd like to utilize the IOMMU core's new behavior,
68 * we could change this to advertise the real page sizes we support.
70 #define AMD_IOMMU_PGSIZES (~0xFFFUL)
72 static DEFINE_RWLOCK(amd_iommu_devtable_lock
);
74 /* A list of preallocated protection domains */
75 static LIST_HEAD(iommu_pd_list
);
76 static DEFINE_SPINLOCK(iommu_pd_list_lock
);
78 /* List of all available dev_data structures */
79 static LIST_HEAD(dev_data_list
);
80 static DEFINE_SPINLOCK(dev_data_list_lock
);
82 LIST_HEAD(ioapic_map
);
86 * Domain for untranslated devices - only allocated
87 * if iommu=pt passed on kernel cmd line.
89 static struct protection_domain
*pt_domain
;
91 static struct iommu_ops amd_iommu_ops
;
93 static ATOMIC_NOTIFIER_HEAD(ppr_notifier
);
94 int amd_iommu_max_glx_val
= -1;
96 static struct dma_map_ops amd_iommu_dma_ops
;
99 * general struct to manage commands send to an IOMMU
105 struct kmem_cache
*amd_iommu_irq_cache
;
107 static void update_domain(struct protection_domain
*domain
);
108 static int __init
alloc_passthrough_domain(void);
110 /****************************************************************************
114 ****************************************************************************/
116 static struct iommu_dev_data
*alloc_dev_data(u16 devid
)
118 struct iommu_dev_data
*dev_data
;
121 dev_data
= kzalloc(sizeof(*dev_data
), GFP_KERNEL
);
125 dev_data
->devid
= devid
;
126 atomic_set(&dev_data
->bind
, 0);
128 spin_lock_irqsave(&dev_data_list_lock
, flags
);
129 list_add_tail(&dev_data
->dev_data_list
, &dev_data_list
);
130 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
135 static void free_dev_data(struct iommu_dev_data
*dev_data
)
139 spin_lock_irqsave(&dev_data_list_lock
, flags
);
140 list_del(&dev_data
->dev_data_list
);
141 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
146 static struct iommu_dev_data
*search_dev_data(u16 devid
)
148 struct iommu_dev_data
*dev_data
;
151 spin_lock_irqsave(&dev_data_list_lock
, flags
);
152 list_for_each_entry(dev_data
, &dev_data_list
, dev_data_list
) {
153 if (dev_data
->devid
== devid
)
160 spin_unlock_irqrestore(&dev_data_list_lock
, flags
);
165 static struct iommu_dev_data
*find_dev_data(u16 devid
)
167 struct iommu_dev_data
*dev_data
;
169 dev_data
= search_dev_data(devid
);
171 if (dev_data
== NULL
)
172 dev_data
= alloc_dev_data(devid
);
177 static inline u16
get_device_id(struct device
*dev
)
179 struct pci_dev
*pdev
= to_pci_dev(dev
);
181 return calc_devid(pdev
->bus
->number
, pdev
->devfn
);
184 static struct iommu_dev_data
*get_dev_data(struct device
*dev
)
186 return dev
->archdata
.iommu
;
189 static bool pci_iommuv2_capable(struct pci_dev
*pdev
)
191 static const int caps
[] = {
194 PCI_EXT_CAP_ID_PASID
,
198 for (i
= 0; i
< 3; ++i
) {
199 pos
= pci_find_ext_capability(pdev
, caps
[i
]);
207 static bool pdev_pri_erratum(struct pci_dev
*pdev
, u32 erratum
)
209 struct iommu_dev_data
*dev_data
;
211 dev_data
= get_dev_data(&pdev
->dev
);
213 return dev_data
->errata
& (1 << erratum
) ? true : false;
217 * In this function the list of preallocated protection domains is traversed to
218 * find the domain for a specific device
220 static struct dma_ops_domain
*find_protection_domain(u16 devid
)
222 struct dma_ops_domain
*entry
, *ret
= NULL
;
224 u16 alias
= amd_iommu_alias_table
[devid
];
226 if (list_empty(&iommu_pd_list
))
229 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
231 list_for_each_entry(entry
, &iommu_pd_list
, list
) {
232 if (entry
->target_dev
== devid
||
233 entry
->target_dev
== alias
) {
239 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
245 * This function checks if the driver got a valid device from the caller to
246 * avoid dereferencing invalid pointers.
248 static bool check_device(struct device
*dev
)
252 if (!dev
|| !dev
->dma_mask
)
255 /* No device or no PCI device */
256 if (dev
->bus
!= &pci_bus_type
)
259 devid
= get_device_id(dev
);
261 /* Out of our scope? */
262 if (devid
> amd_iommu_last_bdf
)
265 if (amd_iommu_rlookup_table
[devid
] == NULL
)
271 static void swap_pci_ref(struct pci_dev
**from
, struct pci_dev
*to
)
277 #define REQ_ACS_FLAGS (PCI_ACS_SV | PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_UF)
279 static struct pci_dev
*get_isolation_root(struct pci_dev
*pdev
)
281 struct pci_dev
*dma_pdev
= pdev
;
283 /* Account for quirked devices */
284 swap_pci_ref(&dma_pdev
, pci_get_dma_source(dma_pdev
));
287 * If it's a multifunction device that does not support our
288 * required ACS flags, add to the same group as function 0.
290 if (dma_pdev
->multifunction
&&
291 !pci_acs_enabled(dma_pdev
, REQ_ACS_FLAGS
))
292 swap_pci_ref(&dma_pdev
,
293 pci_get_slot(dma_pdev
->bus
,
294 PCI_DEVFN(PCI_SLOT(dma_pdev
->devfn
),
298 * Devices on the root bus go through the iommu. If that's not us,
299 * find the next upstream device and test ACS up to the root bus.
300 * Finding the next device may require skipping virtual buses.
302 while (!pci_is_root_bus(dma_pdev
->bus
)) {
303 struct pci_bus
*bus
= dma_pdev
->bus
;
306 if (!pci_is_root_bus(bus
))
312 if (pci_acs_path_enabled(bus
->self
, NULL
, REQ_ACS_FLAGS
))
315 swap_pci_ref(&dma_pdev
, pci_dev_get(bus
->self
));
322 static int init_iommu_group(struct device
*dev
)
324 struct iommu_dev_data
*dev_data
;
325 struct iommu_group
*group
;
326 struct pci_dev
*dma_pdev
= NULL
;
329 group
= iommu_group_get(dev
);
331 iommu_group_put(group
);
335 dev_data
= find_dev_data(get_device_id(dev
));
339 if (dev_data
->alias_data
) {
342 alias
= amd_iommu_alias_table
[dev_data
->devid
];
343 dma_pdev
= pci_get_bus_and_slot(alias
>> 8, alias
& 0xff);
347 dma_pdev
= pci_dev_get(to_pci_dev(dev
));
349 dma_pdev
= get_isolation_root(dma_pdev
);
350 group
= iommu_group_get(&dma_pdev
->dev
);
351 pci_dev_put(dma_pdev
);
353 group
= iommu_group_alloc();
355 return PTR_ERR(group
);
358 ret
= iommu_group_add_device(group
, dev
);
360 iommu_group_put(group
);
365 static int iommu_init_device(struct device
*dev
)
367 struct pci_dev
*pdev
= to_pci_dev(dev
);
368 struct iommu_dev_data
*dev_data
;
372 if (dev
->archdata
.iommu
)
375 dev_data
= find_dev_data(get_device_id(dev
));
379 alias
= amd_iommu_alias_table
[dev_data
->devid
];
380 if (alias
!= dev_data
->devid
) {
381 struct iommu_dev_data
*alias_data
;
383 alias_data
= find_dev_data(alias
);
384 if (alias_data
== NULL
) {
385 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
387 free_dev_data(dev_data
);
390 dev_data
->alias_data
= alias_data
;
393 ret
= init_iommu_group(dev
);
397 if (pci_iommuv2_capable(pdev
)) {
398 struct amd_iommu
*iommu
;
400 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
401 dev_data
->iommu_v2
= iommu
->is_iommu_v2
;
404 dev
->archdata
.iommu
= dev_data
;
409 static void iommu_ignore_device(struct device
*dev
)
413 devid
= get_device_id(dev
);
414 alias
= amd_iommu_alias_table
[devid
];
416 memset(&amd_iommu_dev_table
[devid
], 0, sizeof(struct dev_table_entry
));
417 memset(&amd_iommu_dev_table
[alias
], 0, sizeof(struct dev_table_entry
));
419 amd_iommu_rlookup_table
[devid
] = NULL
;
420 amd_iommu_rlookup_table
[alias
] = NULL
;
423 static void iommu_uninit_device(struct device
*dev
)
425 iommu_group_remove_device(dev
);
428 * Nothing to do here - we keep dev_data around for unplugged devices
429 * and reuse it when the device is re-plugged - not doing so would
430 * introduce a ton of races.
434 void __init
amd_iommu_uninit_devices(void)
436 struct iommu_dev_data
*dev_data
, *n
;
437 struct pci_dev
*pdev
= NULL
;
439 for_each_pci_dev(pdev
) {
441 if (!check_device(&pdev
->dev
))
444 iommu_uninit_device(&pdev
->dev
);
447 /* Free all of our dev_data structures */
448 list_for_each_entry_safe(dev_data
, n
, &dev_data_list
, dev_data_list
)
449 free_dev_data(dev_data
);
452 int __init
amd_iommu_init_devices(void)
454 struct pci_dev
*pdev
= NULL
;
457 for_each_pci_dev(pdev
) {
459 if (!check_device(&pdev
->dev
))
462 ret
= iommu_init_device(&pdev
->dev
);
463 if (ret
== -ENOTSUPP
)
464 iommu_ignore_device(&pdev
->dev
);
473 amd_iommu_uninit_devices();
477 #ifdef CONFIG_AMD_IOMMU_STATS
480 * Initialization code for statistics collection
483 DECLARE_STATS_COUNTER(compl_wait
);
484 DECLARE_STATS_COUNTER(cnt_map_single
);
485 DECLARE_STATS_COUNTER(cnt_unmap_single
);
486 DECLARE_STATS_COUNTER(cnt_map_sg
);
487 DECLARE_STATS_COUNTER(cnt_unmap_sg
);
488 DECLARE_STATS_COUNTER(cnt_alloc_coherent
);
489 DECLARE_STATS_COUNTER(cnt_free_coherent
);
490 DECLARE_STATS_COUNTER(cross_page
);
491 DECLARE_STATS_COUNTER(domain_flush_single
);
492 DECLARE_STATS_COUNTER(domain_flush_all
);
493 DECLARE_STATS_COUNTER(alloced_io_mem
);
494 DECLARE_STATS_COUNTER(total_map_requests
);
495 DECLARE_STATS_COUNTER(complete_ppr
);
496 DECLARE_STATS_COUNTER(invalidate_iotlb
);
497 DECLARE_STATS_COUNTER(invalidate_iotlb_all
);
498 DECLARE_STATS_COUNTER(pri_requests
);
500 static struct dentry
*stats_dir
;
501 static struct dentry
*de_fflush
;
503 static void amd_iommu_stats_add(struct __iommu_counter
*cnt
)
505 if (stats_dir
== NULL
)
508 cnt
->dent
= debugfs_create_u64(cnt
->name
, 0444, stats_dir
,
512 static void amd_iommu_stats_init(void)
514 stats_dir
= debugfs_create_dir("amd-iommu", NULL
);
515 if (stats_dir
== NULL
)
518 de_fflush
= debugfs_create_bool("fullflush", 0444, stats_dir
,
519 &amd_iommu_unmap_flush
);
521 amd_iommu_stats_add(&compl_wait
);
522 amd_iommu_stats_add(&cnt_map_single
);
523 amd_iommu_stats_add(&cnt_unmap_single
);
524 amd_iommu_stats_add(&cnt_map_sg
);
525 amd_iommu_stats_add(&cnt_unmap_sg
);
526 amd_iommu_stats_add(&cnt_alloc_coherent
);
527 amd_iommu_stats_add(&cnt_free_coherent
);
528 amd_iommu_stats_add(&cross_page
);
529 amd_iommu_stats_add(&domain_flush_single
);
530 amd_iommu_stats_add(&domain_flush_all
);
531 amd_iommu_stats_add(&alloced_io_mem
);
532 amd_iommu_stats_add(&total_map_requests
);
533 amd_iommu_stats_add(&complete_ppr
);
534 amd_iommu_stats_add(&invalidate_iotlb
);
535 amd_iommu_stats_add(&invalidate_iotlb_all
);
536 amd_iommu_stats_add(&pri_requests
);
541 /****************************************************************************
543 * Interrupt handling functions
545 ****************************************************************************/
547 static void dump_dte_entry(u16 devid
)
551 for (i
= 0; i
< 4; ++i
)
552 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i
,
553 amd_iommu_dev_table
[devid
].data
[i
]);
556 static void dump_command(unsigned long phys_addr
)
558 struct iommu_cmd
*cmd
= phys_to_virt(phys_addr
);
561 for (i
= 0; i
< 4; ++i
)
562 pr_err("AMD-Vi: CMD[%d]: %08x\n", i
, cmd
->data
[i
]);
565 static void iommu_print_event(struct amd_iommu
*iommu
, void *__evt
)
567 int type
, devid
, domid
, flags
;
568 volatile u32
*event
= __evt
;
573 type
= (event
[1] >> EVENT_TYPE_SHIFT
) & EVENT_TYPE_MASK
;
574 devid
= (event
[0] >> EVENT_DEVID_SHIFT
) & EVENT_DEVID_MASK
;
575 domid
= (event
[1] >> EVENT_DOMID_SHIFT
) & EVENT_DOMID_MASK
;
576 flags
= (event
[1] >> EVENT_FLAGS_SHIFT
) & EVENT_FLAGS_MASK
;
577 address
= (u64
)(((u64
)event
[3]) << 32) | event
[2];
580 /* Did we hit the erratum? */
581 if (++count
== LOOP_TIMEOUT
) {
582 pr_err("AMD-Vi: No event written to event log\n");
589 printk(KERN_ERR
"AMD-Vi: Event logged [");
592 case EVENT_TYPE_ILL_DEV
:
593 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
594 "address=0x%016llx flags=0x%04x]\n",
595 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
597 dump_dte_entry(devid
);
599 case EVENT_TYPE_IO_FAULT
:
600 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
601 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
602 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
603 domid
, address
, flags
);
605 case EVENT_TYPE_DEV_TAB_ERR
:
606 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
607 "address=0x%016llx flags=0x%04x]\n",
608 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
611 case EVENT_TYPE_PAGE_TAB_ERR
:
612 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
613 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
614 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
615 domid
, address
, flags
);
617 case EVENT_TYPE_ILL_CMD
:
618 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address
);
619 dump_command(address
);
621 case EVENT_TYPE_CMD_HARD_ERR
:
622 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
623 "flags=0x%04x]\n", address
, flags
);
625 case EVENT_TYPE_IOTLB_INV_TO
:
626 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
627 "address=0x%016llx]\n",
628 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
631 case EVENT_TYPE_INV_DEV_REQ
:
632 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
633 "address=0x%016llx flags=0x%04x]\n",
634 PCI_BUS(devid
), PCI_SLOT(devid
), PCI_FUNC(devid
),
638 printk(KERN_ERR
"UNKNOWN type=0x%02x]\n", type
);
641 memset(__evt
, 0, 4 * sizeof(u32
));
644 static void iommu_poll_events(struct amd_iommu
*iommu
)
649 spin_lock_irqsave(&iommu
->lock
, flags
);
651 head
= readl(iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
652 tail
= readl(iommu
->mmio_base
+ MMIO_EVT_TAIL_OFFSET
);
654 while (head
!= tail
) {
655 iommu_print_event(iommu
, iommu
->evt_buf
+ head
);
656 head
= (head
+ EVENT_ENTRY_SIZE
) % iommu
->evt_buf_size
;
659 writel(head
, iommu
->mmio_base
+ MMIO_EVT_HEAD_OFFSET
);
661 spin_unlock_irqrestore(&iommu
->lock
, flags
);
664 static void iommu_handle_ppr_entry(struct amd_iommu
*iommu
, u64
*raw
)
666 struct amd_iommu_fault fault
;
668 INC_STATS_COUNTER(pri_requests
);
670 if (PPR_REQ_TYPE(raw
[0]) != PPR_REQ_FAULT
) {
671 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
675 fault
.address
= raw
[1];
676 fault
.pasid
= PPR_PASID(raw
[0]);
677 fault
.device_id
= PPR_DEVID(raw
[0]);
678 fault
.tag
= PPR_TAG(raw
[0]);
679 fault
.flags
= PPR_FLAGS(raw
[0]);
681 atomic_notifier_call_chain(&ppr_notifier
, 0, &fault
);
684 static void iommu_poll_ppr_log(struct amd_iommu
*iommu
)
689 if (iommu
->ppr_log
== NULL
)
692 /* enable ppr interrupts again */
693 writel(MMIO_STATUS_PPR_INT_MASK
, iommu
->mmio_base
+ MMIO_STATUS_OFFSET
);
695 spin_lock_irqsave(&iommu
->lock
, flags
);
697 head
= readl(iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
698 tail
= readl(iommu
->mmio_base
+ MMIO_PPR_TAIL_OFFSET
);
700 while (head
!= tail
) {
705 raw
= (u64
*)(iommu
->ppr_log
+ head
);
708 * Hardware bug: Interrupt may arrive before the entry is
709 * written to memory. If this happens we need to wait for the
712 for (i
= 0; i
< LOOP_TIMEOUT
; ++i
) {
713 if (PPR_REQ_TYPE(raw
[0]) != 0)
718 /* Avoid memcpy function-call overhead */
723 * To detect the hardware bug we need to clear the entry
726 raw
[0] = raw
[1] = 0UL;
728 /* Update head pointer of hardware ring-buffer */
729 head
= (head
+ PPR_ENTRY_SIZE
) % PPR_LOG_SIZE
;
730 writel(head
, iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
733 * Release iommu->lock because ppr-handling might need to
736 spin_unlock_irqrestore(&iommu
->lock
, flags
);
738 /* Handle PPR entry */
739 iommu_handle_ppr_entry(iommu
, entry
);
741 spin_lock_irqsave(&iommu
->lock
, flags
);
743 /* Refresh ring-buffer information */
744 head
= readl(iommu
->mmio_base
+ MMIO_PPR_HEAD_OFFSET
);
745 tail
= readl(iommu
->mmio_base
+ MMIO_PPR_TAIL_OFFSET
);
748 spin_unlock_irqrestore(&iommu
->lock
, flags
);
751 irqreturn_t
amd_iommu_int_thread(int irq
, void *data
)
753 struct amd_iommu
*iommu
;
755 for_each_iommu(iommu
) {
756 iommu_poll_events(iommu
);
757 iommu_poll_ppr_log(iommu
);
763 irqreturn_t
amd_iommu_int_handler(int irq
, void *data
)
765 return IRQ_WAKE_THREAD
;
768 /****************************************************************************
770 * IOMMU command queuing functions
772 ****************************************************************************/
774 static int wait_on_sem(volatile u64
*sem
)
778 while (*sem
== 0 && i
< LOOP_TIMEOUT
) {
783 if (i
== LOOP_TIMEOUT
) {
784 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
791 static void copy_cmd_to_buffer(struct amd_iommu
*iommu
,
792 struct iommu_cmd
*cmd
,
797 target
= iommu
->cmd_buf
+ tail
;
798 tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
800 /* Copy command to buffer */
801 memcpy(target
, cmd
, sizeof(*cmd
));
803 /* Tell the IOMMU about it */
804 writel(tail
, iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
807 static void build_completion_wait(struct iommu_cmd
*cmd
, u64 address
)
809 WARN_ON(address
& 0x7ULL
);
811 memset(cmd
, 0, sizeof(*cmd
));
812 cmd
->data
[0] = lower_32_bits(__pa(address
)) | CMD_COMPL_WAIT_STORE_MASK
;
813 cmd
->data
[1] = upper_32_bits(__pa(address
));
815 CMD_SET_TYPE(cmd
, CMD_COMPL_WAIT
);
818 static void build_inv_dte(struct iommu_cmd
*cmd
, u16 devid
)
820 memset(cmd
, 0, sizeof(*cmd
));
821 cmd
->data
[0] = devid
;
822 CMD_SET_TYPE(cmd
, CMD_INV_DEV_ENTRY
);
825 static void build_inv_iommu_pages(struct iommu_cmd
*cmd
, u64 address
,
826 size_t size
, u16 domid
, int pde
)
831 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
836 * If we have to flush more than one page, flush all
837 * TLB entries for this domain
839 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
843 address
&= PAGE_MASK
;
845 memset(cmd
, 0, sizeof(*cmd
));
846 cmd
->data
[1] |= domid
;
847 cmd
->data
[2] = lower_32_bits(address
);
848 cmd
->data
[3] = upper_32_bits(address
);
849 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
850 if (s
) /* size bit - we flush more than one 4kb page */
851 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
852 if (pde
) /* PDE bit - we want to flush everything, not only the PTEs */
853 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
856 static void build_inv_iotlb_pages(struct iommu_cmd
*cmd
, u16 devid
, int qdep
,
857 u64 address
, size_t size
)
862 pages
= iommu_num_pages(address
, size
, PAGE_SIZE
);
867 * If we have to flush more than one page, flush all
868 * TLB entries for this domain
870 address
= CMD_INV_IOMMU_ALL_PAGES_ADDRESS
;
874 address
&= PAGE_MASK
;
876 memset(cmd
, 0, sizeof(*cmd
));
877 cmd
->data
[0] = devid
;
878 cmd
->data
[0] |= (qdep
& 0xff) << 24;
879 cmd
->data
[1] = devid
;
880 cmd
->data
[2] = lower_32_bits(address
);
881 cmd
->data
[3] = upper_32_bits(address
);
882 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
884 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
887 static void build_inv_iommu_pasid(struct iommu_cmd
*cmd
, u16 domid
, int pasid
,
888 u64 address
, bool size
)
890 memset(cmd
, 0, sizeof(*cmd
));
892 address
&= ~(0xfffULL
);
894 cmd
->data
[0] = pasid
& PASID_MASK
;
895 cmd
->data
[1] = domid
;
896 cmd
->data
[2] = lower_32_bits(address
);
897 cmd
->data
[3] = upper_32_bits(address
);
898 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK
;
899 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_GN_MASK
;
901 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
902 CMD_SET_TYPE(cmd
, CMD_INV_IOMMU_PAGES
);
905 static void build_inv_iotlb_pasid(struct iommu_cmd
*cmd
, u16 devid
, int pasid
,
906 int qdep
, u64 address
, bool size
)
908 memset(cmd
, 0, sizeof(*cmd
));
910 address
&= ~(0xfffULL
);
912 cmd
->data
[0] = devid
;
913 cmd
->data
[0] |= (pasid
& 0xff) << 16;
914 cmd
->data
[0] |= (qdep
& 0xff) << 24;
915 cmd
->data
[1] = devid
;
916 cmd
->data
[1] |= ((pasid
>> 8) & 0xfff) << 16;
917 cmd
->data
[2] = lower_32_bits(address
);
918 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_GN_MASK
;
919 cmd
->data
[3] = upper_32_bits(address
);
921 cmd
->data
[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK
;
922 CMD_SET_TYPE(cmd
, CMD_INV_IOTLB_PAGES
);
925 static void build_complete_ppr(struct iommu_cmd
*cmd
, u16 devid
, int pasid
,
926 int status
, int tag
, bool gn
)
928 memset(cmd
, 0, sizeof(*cmd
));
930 cmd
->data
[0] = devid
;
932 cmd
->data
[1] = pasid
& PASID_MASK
;
933 cmd
->data
[2] = CMD_INV_IOMMU_PAGES_GN_MASK
;
935 cmd
->data
[3] = tag
& 0x1ff;
936 cmd
->data
[3] |= (status
& PPR_STATUS_MASK
) << PPR_STATUS_SHIFT
;
938 CMD_SET_TYPE(cmd
, CMD_COMPLETE_PPR
);
941 static void build_inv_all(struct iommu_cmd
*cmd
)
943 memset(cmd
, 0, sizeof(*cmd
));
944 CMD_SET_TYPE(cmd
, CMD_INV_ALL
);
947 static void build_inv_irt(struct iommu_cmd
*cmd
, u16 devid
)
949 memset(cmd
, 0, sizeof(*cmd
));
950 cmd
->data
[0] = devid
;
951 CMD_SET_TYPE(cmd
, CMD_INV_IRT
);
955 * Writes the command to the IOMMUs command buffer and informs the
956 * hardware about the new command.
958 static int iommu_queue_command_sync(struct amd_iommu
*iommu
,
959 struct iommu_cmd
*cmd
,
962 u32 left
, tail
, head
, next_tail
;
965 WARN_ON(iommu
->cmd_buf_size
& CMD_BUFFER_UNINITIALIZED
);
968 spin_lock_irqsave(&iommu
->lock
, flags
);
970 head
= readl(iommu
->mmio_base
+ MMIO_CMD_HEAD_OFFSET
);
971 tail
= readl(iommu
->mmio_base
+ MMIO_CMD_TAIL_OFFSET
);
972 next_tail
= (tail
+ sizeof(*cmd
)) % iommu
->cmd_buf_size
;
973 left
= (head
- next_tail
) % iommu
->cmd_buf_size
;
976 struct iommu_cmd sync_cmd
;
977 volatile u64 sem
= 0;
980 build_completion_wait(&sync_cmd
, (u64
)&sem
);
981 copy_cmd_to_buffer(iommu
, &sync_cmd
, tail
);
983 spin_unlock_irqrestore(&iommu
->lock
, flags
);
985 if ((ret
= wait_on_sem(&sem
)) != 0)
991 copy_cmd_to_buffer(iommu
, cmd
, tail
);
993 /* We need to sync now to make sure all commands are processed */
994 iommu
->need_sync
= sync
;
996 spin_unlock_irqrestore(&iommu
->lock
, flags
);
1001 static int iommu_queue_command(struct amd_iommu
*iommu
, struct iommu_cmd
*cmd
)
1003 return iommu_queue_command_sync(iommu
, cmd
, true);
1007 * This function queues a completion wait command into the command
1008 * buffer of an IOMMU
1010 static int iommu_completion_wait(struct amd_iommu
*iommu
)
1012 struct iommu_cmd cmd
;
1013 volatile u64 sem
= 0;
1016 if (!iommu
->need_sync
)
1019 build_completion_wait(&cmd
, (u64
)&sem
);
1021 ret
= iommu_queue_command_sync(iommu
, &cmd
, false);
1025 return wait_on_sem(&sem
);
1028 static int iommu_flush_dte(struct amd_iommu
*iommu
, u16 devid
)
1030 struct iommu_cmd cmd
;
1032 build_inv_dte(&cmd
, devid
);
1034 return iommu_queue_command(iommu
, &cmd
);
1037 static void iommu_flush_dte_all(struct amd_iommu
*iommu
)
1041 for (devid
= 0; devid
<= 0xffff; ++devid
)
1042 iommu_flush_dte(iommu
, devid
);
1044 iommu_completion_wait(iommu
);
1048 * This function uses heavy locking and may disable irqs for some time. But
1049 * this is no issue because it is only called during resume.
1051 static void iommu_flush_tlb_all(struct amd_iommu
*iommu
)
1055 for (dom_id
= 0; dom_id
<= 0xffff; ++dom_id
) {
1056 struct iommu_cmd cmd
;
1057 build_inv_iommu_pages(&cmd
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
1059 iommu_queue_command(iommu
, &cmd
);
1062 iommu_completion_wait(iommu
);
1065 static void iommu_flush_all(struct amd_iommu
*iommu
)
1067 struct iommu_cmd cmd
;
1069 build_inv_all(&cmd
);
1071 iommu_queue_command(iommu
, &cmd
);
1072 iommu_completion_wait(iommu
);
1075 static void iommu_flush_irt(struct amd_iommu
*iommu
, u16 devid
)
1077 struct iommu_cmd cmd
;
1079 build_inv_irt(&cmd
, devid
);
1081 iommu_queue_command(iommu
, &cmd
);
1084 static void iommu_flush_irt_all(struct amd_iommu
*iommu
)
1088 for (devid
= 0; devid
<= MAX_DEV_TABLE_ENTRIES
; devid
++)
1089 iommu_flush_irt(iommu
, devid
);
1091 iommu_completion_wait(iommu
);
1094 void iommu_flush_all_caches(struct amd_iommu
*iommu
)
1096 if (iommu_feature(iommu
, FEATURE_IA
)) {
1097 iommu_flush_all(iommu
);
1099 iommu_flush_dte_all(iommu
);
1100 iommu_flush_irt_all(iommu
);
1101 iommu_flush_tlb_all(iommu
);
1106 * Command send function for flushing on-device TLB
1108 static int device_flush_iotlb(struct iommu_dev_data
*dev_data
,
1109 u64 address
, size_t size
)
1111 struct amd_iommu
*iommu
;
1112 struct iommu_cmd cmd
;
1115 qdep
= dev_data
->ats
.qdep
;
1116 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1118 build_inv_iotlb_pages(&cmd
, dev_data
->devid
, qdep
, address
, size
);
1120 return iommu_queue_command(iommu
, &cmd
);
1124 * Command send function for invalidating a device table entry
1126 static int device_flush_dte(struct iommu_dev_data
*dev_data
)
1128 struct amd_iommu
*iommu
;
1131 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
1133 ret
= iommu_flush_dte(iommu
, dev_data
->devid
);
1137 if (dev_data
->ats
.enabled
)
1138 ret
= device_flush_iotlb(dev_data
, 0, ~0UL);
1144 * TLB invalidation function which is called from the mapping functions.
1145 * It invalidates a single PTE if the range to flush is within a single
1146 * page. Otherwise it flushes the whole TLB of the IOMMU.
1148 static void __domain_flush_pages(struct protection_domain
*domain
,
1149 u64 address
, size_t size
, int pde
)
1151 struct iommu_dev_data
*dev_data
;
1152 struct iommu_cmd cmd
;
1155 build_inv_iommu_pages(&cmd
, address
, size
, domain
->id
, pde
);
1157 for (i
= 0; i
< amd_iommus_present
; ++i
) {
1158 if (!domain
->dev_iommu
[i
])
1162 * Devices of this domain are behind this IOMMU
1163 * We need a TLB flush
1165 ret
|= iommu_queue_command(amd_iommus
[i
], &cmd
);
1168 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
1170 if (!dev_data
->ats
.enabled
)
1173 ret
|= device_flush_iotlb(dev_data
, address
, size
);
1179 static void domain_flush_pages(struct protection_domain
*domain
,
1180 u64 address
, size_t size
)
1182 __domain_flush_pages(domain
, address
, size
, 0);
1185 /* Flush the whole IO/TLB for a given protection domain */
1186 static void domain_flush_tlb(struct protection_domain
*domain
)
1188 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 0);
1191 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1192 static void domain_flush_tlb_pde(struct protection_domain
*domain
)
1194 __domain_flush_pages(domain
, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
, 1);
1197 static void domain_flush_complete(struct protection_domain
*domain
)
1201 for (i
= 0; i
< amd_iommus_present
; ++i
) {
1202 if (!domain
->dev_iommu
[i
])
1206 * Devices of this domain are behind this IOMMU
1207 * We need to wait for completion of all commands.
1209 iommu_completion_wait(amd_iommus
[i
]);
1215 * This function flushes the DTEs for all devices in domain
1217 static void domain_flush_devices(struct protection_domain
*domain
)
1219 struct iommu_dev_data
*dev_data
;
1221 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
1222 device_flush_dte(dev_data
);
1225 /****************************************************************************
1227 * The functions below are used the create the page table mappings for
1228 * unity mapped regions.
1230 ****************************************************************************/
1233 * This function is used to add another level to an IO page table. Adding
1234 * another level increases the size of the address space by 9 bits to a size up
1237 static bool increase_address_space(struct protection_domain
*domain
,
1242 if (domain
->mode
== PAGE_MODE_6_LEVEL
)
1243 /* address space already 64 bit large */
1246 pte
= (void *)get_zeroed_page(gfp
);
1250 *pte
= PM_LEVEL_PDE(domain
->mode
,
1251 virt_to_phys(domain
->pt_root
));
1252 domain
->pt_root
= pte
;
1254 domain
->updated
= true;
1259 static u64
*alloc_pte(struct protection_domain
*domain
,
1260 unsigned long address
,
1261 unsigned long page_size
,
1268 BUG_ON(!is_power_of_2(page_size
));
1270 while (address
> PM_LEVEL_SIZE(domain
->mode
))
1271 increase_address_space(domain
, gfp
);
1273 level
= domain
->mode
- 1;
1274 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1275 address
= PAGE_SIZE_ALIGN(address
, page_size
);
1276 end_lvl
= PAGE_SIZE_LEVEL(page_size
);
1278 while (level
> end_lvl
) {
1279 if (!IOMMU_PTE_PRESENT(*pte
)) {
1280 page
= (u64
*)get_zeroed_page(gfp
);
1283 *pte
= PM_LEVEL_PDE(level
, virt_to_phys(page
));
1286 /* No level skipping support yet */
1287 if (PM_PTE_LEVEL(*pte
) != level
)
1292 pte
= IOMMU_PTE_PAGE(*pte
);
1294 if (pte_page
&& level
== end_lvl
)
1297 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1304 * This function checks if there is a PTE for a given dma address. If
1305 * there is one, it returns the pointer to it.
1307 static u64
*fetch_pte(struct protection_domain
*domain
, unsigned long address
)
1312 if (address
> PM_LEVEL_SIZE(domain
->mode
))
1315 level
= domain
->mode
- 1;
1316 pte
= &domain
->pt_root
[PM_LEVEL_INDEX(level
, address
)];
1321 if (!IOMMU_PTE_PRESENT(*pte
))
1325 if (PM_PTE_LEVEL(*pte
) == 0x07) {
1326 unsigned long pte_mask
, __pte
;
1329 * If we have a series of large PTEs, make
1330 * sure to return a pointer to the first one.
1332 pte_mask
= PTE_PAGE_SIZE(*pte
);
1333 pte_mask
= ~((PAGE_SIZE_PTE_COUNT(pte_mask
) << 3) - 1);
1334 __pte
= ((unsigned long)pte
) & pte_mask
;
1336 return (u64
*)__pte
;
1339 /* No level skipping support yet */
1340 if (PM_PTE_LEVEL(*pte
) != level
)
1345 /* Walk to the next level */
1346 pte
= IOMMU_PTE_PAGE(*pte
);
1347 pte
= &pte
[PM_LEVEL_INDEX(level
, address
)];
1354 * Generic mapping functions. It maps a physical address into a DMA
1355 * address space. It allocates the page table pages if necessary.
1356 * In the future it can be extended to a generic mapping function
1357 * supporting all features of AMD IOMMU page tables like level skipping
1358 * and full 64 bit address spaces.
1360 static int iommu_map_page(struct protection_domain
*dom
,
1361 unsigned long bus_addr
,
1362 unsigned long phys_addr
,
1364 unsigned long page_size
)
1369 if (!(prot
& IOMMU_PROT_MASK
))
1372 bus_addr
= PAGE_ALIGN(bus_addr
);
1373 phys_addr
= PAGE_ALIGN(phys_addr
);
1374 count
= PAGE_SIZE_PTE_COUNT(page_size
);
1375 pte
= alloc_pte(dom
, bus_addr
, page_size
, NULL
, GFP_KERNEL
);
1377 for (i
= 0; i
< count
; ++i
)
1378 if (IOMMU_PTE_PRESENT(pte
[i
]))
1381 if (page_size
> PAGE_SIZE
) {
1382 __pte
= PAGE_SIZE_PTE(phys_addr
, page_size
);
1383 __pte
|= PM_LEVEL_ENC(7) | IOMMU_PTE_P
| IOMMU_PTE_FC
;
1385 __pte
= phys_addr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
1387 if (prot
& IOMMU_PROT_IR
)
1388 __pte
|= IOMMU_PTE_IR
;
1389 if (prot
& IOMMU_PROT_IW
)
1390 __pte
|= IOMMU_PTE_IW
;
1392 for (i
= 0; i
< count
; ++i
)
1400 static unsigned long iommu_unmap_page(struct protection_domain
*dom
,
1401 unsigned long bus_addr
,
1402 unsigned long page_size
)
1404 unsigned long long unmap_size
, unmapped
;
1407 BUG_ON(!is_power_of_2(page_size
));
1411 while (unmapped
< page_size
) {
1413 pte
= fetch_pte(dom
, bus_addr
);
1417 * No PTE for this address
1418 * move forward in 4kb steps
1420 unmap_size
= PAGE_SIZE
;
1421 } else if (PM_PTE_LEVEL(*pte
) == 0) {
1422 /* 4kb PTE found for this address */
1423 unmap_size
= PAGE_SIZE
;
1428 /* Large PTE found which maps this address */
1429 unmap_size
= PTE_PAGE_SIZE(*pte
);
1430 count
= PAGE_SIZE_PTE_COUNT(unmap_size
);
1431 for (i
= 0; i
< count
; i
++)
1435 bus_addr
= (bus_addr
& ~(unmap_size
- 1)) + unmap_size
;
1436 unmapped
+= unmap_size
;
1439 BUG_ON(!is_power_of_2(unmapped
));
1445 * This function checks if a specific unity mapping entry is needed for
1446 * this specific IOMMU.
1448 static int iommu_for_unity_map(struct amd_iommu
*iommu
,
1449 struct unity_map_entry
*entry
)
1453 for (i
= entry
->devid_start
; i
<= entry
->devid_end
; ++i
) {
1454 bdf
= amd_iommu_alias_table
[i
];
1455 if (amd_iommu_rlookup_table
[bdf
] == iommu
)
1463 * This function actually applies the mapping to the page table of the
1466 static int dma_ops_unity_map(struct dma_ops_domain
*dma_dom
,
1467 struct unity_map_entry
*e
)
1472 for (addr
= e
->address_start
; addr
< e
->address_end
;
1473 addr
+= PAGE_SIZE
) {
1474 ret
= iommu_map_page(&dma_dom
->domain
, addr
, addr
, e
->prot
,
1479 * if unity mapping is in aperture range mark the page
1480 * as allocated in the aperture
1482 if (addr
< dma_dom
->aperture_size
)
1483 __set_bit(addr
>> PAGE_SHIFT
,
1484 dma_dom
->aperture
[0]->bitmap
);
1491 * Init the unity mappings for a specific IOMMU in the system
1493 * Basically iterates over all unity mapping entries and applies them to
1494 * the default domain DMA of that IOMMU if necessary.
1496 static int iommu_init_unity_mappings(struct amd_iommu
*iommu
)
1498 struct unity_map_entry
*entry
;
1501 list_for_each_entry(entry
, &amd_iommu_unity_map
, list
) {
1502 if (!iommu_for_unity_map(iommu
, entry
))
1504 ret
= dma_ops_unity_map(iommu
->default_dom
, entry
);
1513 * Inits the unity mappings required for a specific device
1515 static int init_unity_mappings_for_device(struct dma_ops_domain
*dma_dom
,
1518 struct unity_map_entry
*e
;
1521 list_for_each_entry(e
, &amd_iommu_unity_map
, list
) {
1522 if (!(devid
>= e
->devid_start
&& devid
<= e
->devid_end
))
1524 ret
= dma_ops_unity_map(dma_dom
, e
);
1532 /****************************************************************************
1534 * The next functions belong to the address allocator for the dma_ops
1535 * interface functions. They work like the allocators in the other IOMMU
1536 * drivers. Its basically a bitmap which marks the allocated pages in
1537 * the aperture. Maybe it could be enhanced in the future to a more
1538 * efficient allocator.
1540 ****************************************************************************/
1543 * The address allocator core functions.
1545 * called with domain->lock held
1549 * Used to reserve address ranges in the aperture (e.g. for exclusion
1552 static void dma_ops_reserve_addresses(struct dma_ops_domain
*dom
,
1553 unsigned long start_page
,
1556 unsigned int i
, last_page
= dom
->aperture_size
>> PAGE_SHIFT
;
1558 if (start_page
+ pages
> last_page
)
1559 pages
= last_page
- start_page
;
1561 for (i
= start_page
; i
< start_page
+ pages
; ++i
) {
1562 int index
= i
/ APERTURE_RANGE_PAGES
;
1563 int page
= i
% APERTURE_RANGE_PAGES
;
1564 __set_bit(page
, dom
->aperture
[index
]->bitmap
);
1569 * This function is used to add a new aperture range to an existing
1570 * aperture in case of dma_ops domain allocation or address allocation
1573 static int alloc_new_range(struct dma_ops_domain
*dma_dom
,
1574 bool populate
, gfp_t gfp
)
1576 int index
= dma_dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1577 struct amd_iommu
*iommu
;
1578 unsigned long i
, old_size
;
1580 #ifdef CONFIG_IOMMU_STRESS
1584 if (index
>= APERTURE_MAX_RANGES
)
1587 dma_dom
->aperture
[index
] = kzalloc(sizeof(struct aperture_range
), gfp
);
1588 if (!dma_dom
->aperture
[index
])
1591 dma_dom
->aperture
[index
]->bitmap
= (void *)get_zeroed_page(gfp
);
1592 if (!dma_dom
->aperture
[index
]->bitmap
)
1595 dma_dom
->aperture
[index
]->offset
= dma_dom
->aperture_size
;
1598 unsigned long address
= dma_dom
->aperture_size
;
1599 int i
, num_ptes
= APERTURE_RANGE_PAGES
/ 512;
1600 u64
*pte
, *pte_page
;
1602 for (i
= 0; i
< num_ptes
; ++i
) {
1603 pte
= alloc_pte(&dma_dom
->domain
, address
, PAGE_SIZE
,
1608 dma_dom
->aperture
[index
]->pte_pages
[i
] = pte_page
;
1610 address
+= APERTURE_RANGE_SIZE
/ 64;
1614 old_size
= dma_dom
->aperture_size
;
1615 dma_dom
->aperture_size
+= APERTURE_RANGE_SIZE
;
1617 /* Reserve address range used for MSI messages */
1618 if (old_size
< MSI_ADDR_BASE_LO
&&
1619 dma_dom
->aperture_size
> MSI_ADDR_BASE_LO
) {
1620 unsigned long spage
;
1623 pages
= iommu_num_pages(MSI_ADDR_BASE_LO
, 0x10000, PAGE_SIZE
);
1624 spage
= MSI_ADDR_BASE_LO
>> PAGE_SHIFT
;
1626 dma_ops_reserve_addresses(dma_dom
, spage
, pages
);
1629 /* Initialize the exclusion range if necessary */
1630 for_each_iommu(iommu
) {
1631 if (iommu
->exclusion_start
&&
1632 iommu
->exclusion_start
>= dma_dom
->aperture
[index
]->offset
1633 && iommu
->exclusion_start
< dma_dom
->aperture_size
) {
1634 unsigned long startpage
;
1635 int pages
= iommu_num_pages(iommu
->exclusion_start
,
1636 iommu
->exclusion_length
,
1638 startpage
= iommu
->exclusion_start
>> PAGE_SHIFT
;
1639 dma_ops_reserve_addresses(dma_dom
, startpage
, pages
);
1644 * Check for areas already mapped as present in the new aperture
1645 * range and mark those pages as reserved in the allocator. Such
1646 * mappings may already exist as a result of requested unity
1647 * mappings for devices.
1649 for (i
= dma_dom
->aperture
[index
]->offset
;
1650 i
< dma_dom
->aperture_size
;
1652 u64
*pte
= fetch_pte(&dma_dom
->domain
, i
);
1653 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
1656 dma_ops_reserve_addresses(dma_dom
, i
>> PAGE_SHIFT
, 1);
1659 update_domain(&dma_dom
->domain
);
1664 update_domain(&dma_dom
->domain
);
1666 free_page((unsigned long)dma_dom
->aperture
[index
]->bitmap
);
1668 kfree(dma_dom
->aperture
[index
]);
1669 dma_dom
->aperture
[index
] = NULL
;
1674 static unsigned long dma_ops_area_alloc(struct device
*dev
,
1675 struct dma_ops_domain
*dom
,
1677 unsigned long align_mask
,
1679 unsigned long start
)
1681 unsigned long next_bit
= dom
->next_address
% APERTURE_RANGE_SIZE
;
1682 int max_index
= dom
->aperture_size
>> APERTURE_RANGE_SHIFT
;
1683 int i
= start
>> APERTURE_RANGE_SHIFT
;
1684 unsigned long boundary_size
;
1685 unsigned long address
= -1;
1686 unsigned long limit
;
1688 next_bit
>>= PAGE_SHIFT
;
1690 boundary_size
= ALIGN(dma_get_seg_boundary(dev
) + 1,
1691 PAGE_SIZE
) >> PAGE_SHIFT
;
1693 for (;i
< max_index
; ++i
) {
1694 unsigned long offset
= dom
->aperture
[i
]->offset
>> PAGE_SHIFT
;
1696 if (dom
->aperture
[i
]->offset
>= dma_mask
)
1699 limit
= iommu_device_max_index(APERTURE_RANGE_PAGES
, offset
,
1700 dma_mask
>> PAGE_SHIFT
);
1702 address
= iommu_area_alloc(dom
->aperture
[i
]->bitmap
,
1703 limit
, next_bit
, pages
, 0,
1704 boundary_size
, align_mask
);
1705 if (address
!= -1) {
1706 address
= dom
->aperture
[i
]->offset
+
1707 (address
<< PAGE_SHIFT
);
1708 dom
->next_address
= address
+ (pages
<< PAGE_SHIFT
);
1718 static unsigned long dma_ops_alloc_addresses(struct device
*dev
,
1719 struct dma_ops_domain
*dom
,
1721 unsigned long align_mask
,
1724 unsigned long address
;
1726 #ifdef CONFIG_IOMMU_STRESS
1727 dom
->next_address
= 0;
1728 dom
->need_flush
= true;
1731 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1732 dma_mask
, dom
->next_address
);
1734 if (address
== -1) {
1735 dom
->next_address
= 0;
1736 address
= dma_ops_area_alloc(dev
, dom
, pages
, align_mask
,
1738 dom
->need_flush
= true;
1741 if (unlikely(address
== -1))
1742 address
= DMA_ERROR_CODE
;
1744 WARN_ON((address
+ (PAGE_SIZE
*pages
)) > dom
->aperture_size
);
1750 * The address free function.
1752 * called with domain->lock held
1754 static void dma_ops_free_addresses(struct dma_ops_domain
*dom
,
1755 unsigned long address
,
1758 unsigned i
= address
>> APERTURE_RANGE_SHIFT
;
1759 struct aperture_range
*range
= dom
->aperture
[i
];
1761 BUG_ON(i
>= APERTURE_MAX_RANGES
|| range
== NULL
);
1763 #ifdef CONFIG_IOMMU_STRESS
1768 if (address
>= dom
->next_address
)
1769 dom
->need_flush
= true;
1771 address
= (address
% APERTURE_RANGE_SIZE
) >> PAGE_SHIFT
;
1773 bitmap_clear(range
->bitmap
, address
, pages
);
1777 /****************************************************************************
1779 * The next functions belong to the domain allocation. A domain is
1780 * allocated for every IOMMU as the default domain. If device isolation
1781 * is enabled, every device get its own domain. The most important thing
1782 * about domains is the page table mapping the DMA address space they
1785 ****************************************************************************/
1788 * This function adds a protection domain to the global protection domain list
1790 static void add_domain_to_list(struct protection_domain
*domain
)
1792 unsigned long flags
;
1794 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1795 list_add(&domain
->list
, &amd_iommu_pd_list
);
1796 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1800 * This function removes a protection domain to the global
1801 * protection domain list
1803 static void del_domain_from_list(struct protection_domain
*domain
)
1805 unsigned long flags
;
1807 spin_lock_irqsave(&amd_iommu_pd_lock
, flags
);
1808 list_del(&domain
->list
);
1809 spin_unlock_irqrestore(&amd_iommu_pd_lock
, flags
);
1812 static u16
domain_id_alloc(void)
1814 unsigned long flags
;
1817 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1818 id
= find_first_zero_bit(amd_iommu_pd_alloc_bitmap
, MAX_DOMAIN_ID
);
1820 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1821 __set_bit(id
, amd_iommu_pd_alloc_bitmap
);
1824 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1829 static void domain_id_free(int id
)
1831 unsigned long flags
;
1833 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
1834 if (id
> 0 && id
< MAX_DOMAIN_ID
)
1835 __clear_bit(id
, amd_iommu_pd_alloc_bitmap
);
1836 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
1839 static void free_pagetable(struct protection_domain
*domain
)
1844 p1
= domain
->pt_root
;
1849 for (i
= 0; i
< 512; ++i
) {
1850 if (!IOMMU_PTE_PRESENT(p1
[i
]))
1853 p2
= IOMMU_PTE_PAGE(p1
[i
]);
1854 for (j
= 0; j
< 512; ++j
) {
1855 if (!IOMMU_PTE_PRESENT(p2
[j
]))
1857 p3
= IOMMU_PTE_PAGE(p2
[j
]);
1858 free_page((unsigned long)p3
);
1861 free_page((unsigned long)p2
);
1864 free_page((unsigned long)p1
);
1866 domain
->pt_root
= NULL
;
1869 static void free_gcr3_tbl_level1(u64
*tbl
)
1874 for (i
= 0; i
< 512; ++i
) {
1875 if (!(tbl
[i
] & GCR3_VALID
))
1878 ptr
= __va(tbl
[i
] & PAGE_MASK
);
1880 free_page((unsigned long)ptr
);
1884 static void free_gcr3_tbl_level2(u64
*tbl
)
1889 for (i
= 0; i
< 512; ++i
) {
1890 if (!(tbl
[i
] & GCR3_VALID
))
1893 ptr
= __va(tbl
[i
] & PAGE_MASK
);
1895 free_gcr3_tbl_level1(ptr
);
1899 static void free_gcr3_table(struct protection_domain
*domain
)
1901 if (domain
->glx
== 2)
1902 free_gcr3_tbl_level2(domain
->gcr3_tbl
);
1903 else if (domain
->glx
== 1)
1904 free_gcr3_tbl_level1(domain
->gcr3_tbl
);
1905 else if (domain
->glx
!= 0)
1908 free_page((unsigned long)domain
->gcr3_tbl
);
1912 * Free a domain, only used if something went wrong in the
1913 * allocation path and we need to free an already allocated page table
1915 static void dma_ops_domain_free(struct dma_ops_domain
*dom
)
1922 del_domain_from_list(&dom
->domain
);
1924 free_pagetable(&dom
->domain
);
1926 for (i
= 0; i
< APERTURE_MAX_RANGES
; ++i
) {
1927 if (!dom
->aperture
[i
])
1929 free_page((unsigned long)dom
->aperture
[i
]->bitmap
);
1930 kfree(dom
->aperture
[i
]);
1937 * Allocates a new protection domain usable for the dma_ops functions.
1938 * It also initializes the page table and the address allocator data
1939 * structures required for the dma_ops interface
1941 static struct dma_ops_domain
*dma_ops_domain_alloc(void)
1943 struct dma_ops_domain
*dma_dom
;
1945 dma_dom
= kzalloc(sizeof(struct dma_ops_domain
), GFP_KERNEL
);
1949 spin_lock_init(&dma_dom
->domain
.lock
);
1951 dma_dom
->domain
.id
= domain_id_alloc();
1952 if (dma_dom
->domain
.id
== 0)
1954 INIT_LIST_HEAD(&dma_dom
->domain
.dev_list
);
1955 dma_dom
->domain
.mode
= PAGE_MODE_2_LEVEL
;
1956 dma_dom
->domain
.pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
1957 dma_dom
->domain
.flags
= PD_DMA_OPS_MASK
;
1958 dma_dom
->domain
.priv
= dma_dom
;
1959 if (!dma_dom
->domain
.pt_root
)
1962 dma_dom
->need_flush
= false;
1963 dma_dom
->target_dev
= 0xffff;
1965 add_domain_to_list(&dma_dom
->domain
);
1967 if (alloc_new_range(dma_dom
, true, GFP_KERNEL
))
1971 * mark the first page as allocated so we never return 0 as
1972 * a valid dma-address. So we can use 0 as error value
1974 dma_dom
->aperture
[0]->bitmap
[0] = 1;
1975 dma_dom
->next_address
= 0;
1981 dma_ops_domain_free(dma_dom
);
1987 * little helper function to check whether a given protection domain is a
1990 static bool dma_ops_domain(struct protection_domain
*domain
)
1992 return domain
->flags
& PD_DMA_OPS_MASK
;
1995 static void set_dte_entry(u16 devid
, struct protection_domain
*domain
, bool ats
)
2000 if (domain
->mode
!= PAGE_MODE_NONE
)
2001 pte_root
= virt_to_phys(domain
->pt_root
);
2003 pte_root
|= (domain
->mode
& DEV_ENTRY_MODE_MASK
)
2004 << DEV_ENTRY_MODE_SHIFT
;
2005 pte_root
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
| IOMMU_PTE_P
| IOMMU_PTE_TV
;
2007 flags
= amd_iommu_dev_table
[devid
].data
[1];
2010 flags
|= DTE_FLAG_IOTLB
;
2012 if (domain
->flags
& PD_IOMMUV2_MASK
) {
2013 u64 gcr3
= __pa(domain
->gcr3_tbl
);
2014 u64 glx
= domain
->glx
;
2017 pte_root
|= DTE_FLAG_GV
;
2018 pte_root
|= (glx
& DTE_GLX_MASK
) << DTE_GLX_SHIFT
;
2020 /* First mask out possible old values for GCR3 table */
2021 tmp
= DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B
;
2024 tmp
= DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C
;
2027 /* Encode GCR3 table into DTE */
2028 tmp
= DTE_GCR3_VAL_A(gcr3
) << DTE_GCR3_SHIFT_A
;
2031 tmp
= DTE_GCR3_VAL_B(gcr3
) << DTE_GCR3_SHIFT_B
;
2034 tmp
= DTE_GCR3_VAL_C(gcr3
) << DTE_GCR3_SHIFT_C
;
2038 flags
&= ~(0xffffUL
);
2039 flags
|= domain
->id
;
2041 amd_iommu_dev_table
[devid
].data
[1] = flags
;
2042 amd_iommu_dev_table
[devid
].data
[0] = pte_root
;
2045 static void clear_dte_entry(u16 devid
)
2047 /* remove entry from the device table seen by the hardware */
2048 amd_iommu_dev_table
[devid
].data
[0] = IOMMU_PTE_P
| IOMMU_PTE_TV
;
2049 amd_iommu_dev_table
[devid
].data
[1] = 0;
2051 amd_iommu_apply_erratum_63(devid
);
2054 static void do_attach(struct iommu_dev_data
*dev_data
,
2055 struct protection_domain
*domain
)
2057 struct amd_iommu
*iommu
;
2060 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
2061 ats
= dev_data
->ats
.enabled
;
2063 /* Update data structures */
2064 dev_data
->domain
= domain
;
2065 list_add(&dev_data
->list
, &domain
->dev_list
);
2066 set_dte_entry(dev_data
->devid
, domain
, ats
);
2068 /* Do reference counting */
2069 domain
->dev_iommu
[iommu
->index
] += 1;
2070 domain
->dev_cnt
+= 1;
2072 /* Flush the DTE entry */
2073 device_flush_dte(dev_data
);
2076 static void do_detach(struct iommu_dev_data
*dev_data
)
2078 struct amd_iommu
*iommu
;
2080 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
2082 /* decrease reference counters */
2083 dev_data
->domain
->dev_iommu
[iommu
->index
] -= 1;
2084 dev_data
->domain
->dev_cnt
-= 1;
2086 /* Update data structures */
2087 dev_data
->domain
= NULL
;
2088 list_del(&dev_data
->list
);
2089 clear_dte_entry(dev_data
->devid
);
2091 /* Flush the DTE entry */
2092 device_flush_dte(dev_data
);
2096 * If a device is not yet associated with a domain, this function does
2097 * assigns it visible for the hardware
2099 static int __attach_device(struct iommu_dev_data
*dev_data
,
2100 struct protection_domain
*domain
)
2105 spin_lock(&domain
->lock
);
2107 if (dev_data
->alias_data
!= NULL
) {
2108 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
2110 /* Some sanity checks */
2112 if (alias_data
->domain
!= NULL
&&
2113 alias_data
->domain
!= domain
)
2116 if (dev_data
->domain
!= NULL
&&
2117 dev_data
->domain
!= domain
)
2120 /* Do real assignment */
2121 if (alias_data
->domain
== NULL
)
2122 do_attach(alias_data
, domain
);
2124 atomic_inc(&alias_data
->bind
);
2127 if (dev_data
->domain
== NULL
)
2128 do_attach(dev_data
, domain
);
2130 atomic_inc(&dev_data
->bind
);
2137 spin_unlock(&domain
->lock
);
2143 static void pdev_iommuv2_disable(struct pci_dev
*pdev
)
2145 pci_disable_ats(pdev
);
2146 pci_disable_pri(pdev
);
2147 pci_disable_pasid(pdev
);
2150 /* FIXME: Change generic reset-function to do the same */
2151 static int pri_reset_while_enabled(struct pci_dev
*pdev
)
2156 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
2160 pci_read_config_word(pdev
, pos
+ PCI_PRI_CTRL
, &control
);
2161 control
|= PCI_PRI_CTRL_RESET
;
2162 pci_write_config_word(pdev
, pos
+ PCI_PRI_CTRL
, control
);
2167 static int pdev_iommuv2_enable(struct pci_dev
*pdev
)
2172 /* FIXME: Hardcode number of outstanding requests for now */
2174 if (pdev_pri_erratum(pdev
, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE
))
2176 reset_enable
= pdev_pri_erratum(pdev
, AMD_PRI_DEV_ERRATUM_ENABLE_RESET
);
2178 /* Only allow access to user-accessible pages */
2179 ret
= pci_enable_pasid(pdev
, 0);
2183 /* First reset the PRI state of the device */
2184 ret
= pci_reset_pri(pdev
);
2189 ret
= pci_enable_pri(pdev
, reqs
);
2194 ret
= pri_reset_while_enabled(pdev
);
2199 ret
= pci_enable_ats(pdev
, PAGE_SHIFT
);
2206 pci_disable_pri(pdev
);
2207 pci_disable_pasid(pdev
);
2212 /* FIXME: Move this to PCI code */
2213 #define PCI_PRI_TLP_OFF (1 << 15)
2215 static bool pci_pri_tlp_required(struct pci_dev
*pdev
)
2220 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
2224 pci_read_config_word(pdev
, pos
+ PCI_PRI_STATUS
, &status
);
2226 return (status
& PCI_PRI_TLP_OFF
) ? true : false;
2230 * If a device is not yet associated with a domain, this function
2231 * assigns it visible for the hardware
2233 static int attach_device(struct device
*dev
,
2234 struct protection_domain
*domain
)
2236 struct pci_dev
*pdev
= to_pci_dev(dev
);
2237 struct iommu_dev_data
*dev_data
;
2238 unsigned long flags
;
2241 dev_data
= get_dev_data(dev
);
2243 if (domain
->flags
& PD_IOMMUV2_MASK
) {
2244 if (!dev_data
->iommu_v2
|| !dev_data
->passthrough
)
2247 if (pdev_iommuv2_enable(pdev
) != 0)
2250 dev_data
->ats
.enabled
= true;
2251 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
2252 dev_data
->pri_tlp
= pci_pri_tlp_required(pdev
);
2253 } else if (amd_iommu_iotlb_sup
&&
2254 pci_enable_ats(pdev
, PAGE_SHIFT
) == 0) {
2255 dev_data
->ats
.enabled
= true;
2256 dev_data
->ats
.qdep
= pci_ats_queue_depth(pdev
);
2259 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2260 ret
= __attach_device(dev_data
, domain
);
2261 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2264 * We might boot into a crash-kernel here. The crashed kernel
2265 * left the caches in the IOMMU dirty. So we have to flush
2266 * here to evict all dirty stuff.
2268 domain_flush_tlb_pde(domain
);
2274 * Removes a device from a protection domain (unlocked)
2276 static void __detach_device(struct iommu_dev_data
*dev_data
)
2278 struct protection_domain
*domain
;
2279 unsigned long flags
;
2281 BUG_ON(!dev_data
->domain
);
2283 domain
= dev_data
->domain
;
2285 spin_lock_irqsave(&domain
->lock
, flags
);
2287 if (dev_data
->alias_data
!= NULL
) {
2288 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
2290 if (atomic_dec_and_test(&alias_data
->bind
))
2291 do_detach(alias_data
);
2294 if (atomic_dec_and_test(&dev_data
->bind
))
2295 do_detach(dev_data
);
2297 spin_unlock_irqrestore(&domain
->lock
, flags
);
2300 * If we run in passthrough mode the device must be assigned to the
2301 * passthrough domain if it is detached from any other domain.
2302 * Make sure we can deassign from the pt_domain itself.
2304 if (dev_data
->passthrough
&&
2305 (dev_data
->domain
== NULL
&& domain
!= pt_domain
))
2306 __attach_device(dev_data
, pt_domain
);
2310 * Removes a device from a protection domain (with devtable_lock held)
2312 static void detach_device(struct device
*dev
)
2314 struct protection_domain
*domain
;
2315 struct iommu_dev_data
*dev_data
;
2316 unsigned long flags
;
2318 dev_data
= get_dev_data(dev
);
2319 domain
= dev_data
->domain
;
2321 /* lock device table */
2322 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2323 __detach_device(dev_data
);
2324 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2326 if (domain
->flags
& PD_IOMMUV2_MASK
)
2327 pdev_iommuv2_disable(to_pci_dev(dev
));
2328 else if (dev_data
->ats
.enabled
)
2329 pci_disable_ats(to_pci_dev(dev
));
2331 dev_data
->ats
.enabled
= false;
2335 * Find out the protection domain structure for a given PCI device. This
2336 * will give us the pointer to the page table root for example.
2338 static struct protection_domain
*domain_for_device(struct device
*dev
)
2340 struct iommu_dev_data
*dev_data
;
2341 struct protection_domain
*dom
= NULL
;
2342 unsigned long flags
;
2344 dev_data
= get_dev_data(dev
);
2346 if (dev_data
->domain
)
2347 return dev_data
->domain
;
2349 if (dev_data
->alias_data
!= NULL
) {
2350 struct iommu_dev_data
*alias_data
= dev_data
->alias_data
;
2352 read_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
2353 if (alias_data
->domain
!= NULL
) {
2354 __attach_device(dev_data
, alias_data
->domain
);
2355 dom
= alias_data
->domain
;
2357 read_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
2363 static int device_change_notifier(struct notifier_block
*nb
,
2364 unsigned long action
, void *data
)
2366 struct dma_ops_domain
*dma_domain
;
2367 struct protection_domain
*domain
;
2368 struct iommu_dev_data
*dev_data
;
2369 struct device
*dev
= data
;
2370 struct amd_iommu
*iommu
;
2371 unsigned long flags
;
2374 if (!check_device(dev
))
2377 devid
= get_device_id(dev
);
2378 iommu
= amd_iommu_rlookup_table
[devid
];
2379 dev_data
= get_dev_data(dev
);
2382 case BUS_NOTIFY_UNBOUND_DRIVER
:
2384 domain
= domain_for_device(dev
);
2388 if (dev_data
->passthrough
)
2392 case BUS_NOTIFY_ADD_DEVICE
:
2394 iommu_init_device(dev
);
2397 * dev_data is still NULL and
2398 * got initialized in iommu_init_device
2400 dev_data
= get_dev_data(dev
);
2402 if (iommu_pass_through
|| dev_data
->iommu_v2
) {
2403 dev_data
->passthrough
= true;
2404 attach_device(dev
, pt_domain
);
2408 domain
= domain_for_device(dev
);
2410 /* allocate a protection domain if a device is added */
2411 dma_domain
= find_protection_domain(devid
);
2414 dma_domain
= dma_ops_domain_alloc();
2417 dma_domain
->target_dev
= devid
;
2419 spin_lock_irqsave(&iommu_pd_list_lock
, flags
);
2420 list_add_tail(&dma_domain
->list
, &iommu_pd_list
);
2421 spin_unlock_irqrestore(&iommu_pd_list_lock
, flags
);
2423 dev_data
= get_dev_data(dev
);
2425 dev
->archdata
.dma_ops
= &amd_iommu_dma_ops
;
2428 case BUS_NOTIFY_DEL_DEVICE
:
2430 iommu_uninit_device(dev
);
2436 iommu_completion_wait(iommu
);
2442 static struct notifier_block device_nb
= {
2443 .notifier_call
= device_change_notifier
,
2446 void amd_iommu_init_notifier(void)
2448 bus_register_notifier(&pci_bus_type
, &device_nb
);
2451 /*****************************************************************************
2453 * The next functions belong to the dma_ops mapping/unmapping code.
2455 *****************************************************************************/
2458 * In the dma_ops path we only have the struct device. This function
2459 * finds the corresponding IOMMU, the protection domain and the
2460 * requestor id for a given device.
2461 * If the device is not yet associated with a domain this is also done
2464 static struct protection_domain
*get_domain(struct device
*dev
)
2466 struct protection_domain
*domain
;
2467 struct dma_ops_domain
*dma_dom
;
2468 u16 devid
= get_device_id(dev
);
2470 if (!check_device(dev
))
2471 return ERR_PTR(-EINVAL
);
2473 domain
= domain_for_device(dev
);
2474 if (domain
!= NULL
&& !dma_ops_domain(domain
))
2475 return ERR_PTR(-EBUSY
);
2480 /* Device not bound yet - bind it */
2481 dma_dom
= find_protection_domain(devid
);
2483 dma_dom
= amd_iommu_rlookup_table
[devid
]->default_dom
;
2484 attach_device(dev
, &dma_dom
->domain
);
2485 DUMP_printk("Using protection domain %d for device %s\n",
2486 dma_dom
->domain
.id
, dev_name(dev
));
2488 return &dma_dom
->domain
;
2491 static void update_device_table(struct protection_domain
*domain
)
2493 struct iommu_dev_data
*dev_data
;
2495 list_for_each_entry(dev_data
, &domain
->dev_list
, list
)
2496 set_dte_entry(dev_data
->devid
, domain
, dev_data
->ats
.enabled
);
2499 static void update_domain(struct protection_domain
*domain
)
2501 if (!domain
->updated
)
2504 update_device_table(domain
);
2506 domain_flush_devices(domain
);
2507 domain_flush_tlb_pde(domain
);
2509 domain
->updated
= false;
2513 * This function fetches the PTE for a given address in the aperture
2515 static u64
* dma_ops_get_pte(struct dma_ops_domain
*dom
,
2516 unsigned long address
)
2518 struct aperture_range
*aperture
;
2519 u64
*pte
, *pte_page
;
2521 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
2525 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
2527 pte
= alloc_pte(&dom
->domain
, address
, PAGE_SIZE
, &pte_page
,
2529 aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)] = pte_page
;
2531 pte
+= PM_LEVEL_INDEX(0, address
);
2533 update_domain(&dom
->domain
);
2539 * This is the generic map function. It maps one 4kb page at paddr to
2540 * the given address in the DMA address space for the domain.
2542 static dma_addr_t
dma_ops_domain_map(struct dma_ops_domain
*dom
,
2543 unsigned long address
,
2549 WARN_ON(address
> dom
->aperture_size
);
2553 pte
= dma_ops_get_pte(dom
, address
);
2555 return DMA_ERROR_CODE
;
2557 __pte
= paddr
| IOMMU_PTE_P
| IOMMU_PTE_FC
;
2559 if (direction
== DMA_TO_DEVICE
)
2560 __pte
|= IOMMU_PTE_IR
;
2561 else if (direction
== DMA_FROM_DEVICE
)
2562 __pte
|= IOMMU_PTE_IW
;
2563 else if (direction
== DMA_BIDIRECTIONAL
)
2564 __pte
|= IOMMU_PTE_IR
| IOMMU_PTE_IW
;
2570 return (dma_addr_t
)address
;
2574 * The generic unmapping function for on page in the DMA address space.
2576 static void dma_ops_domain_unmap(struct dma_ops_domain
*dom
,
2577 unsigned long address
)
2579 struct aperture_range
*aperture
;
2582 if (address
>= dom
->aperture_size
)
2585 aperture
= dom
->aperture
[APERTURE_RANGE_INDEX(address
)];
2589 pte
= aperture
->pte_pages
[APERTURE_PAGE_INDEX(address
)];
2593 pte
+= PM_LEVEL_INDEX(0, address
);
2601 * This function contains common code for mapping of a physically
2602 * contiguous memory region into DMA address space. It is used by all
2603 * mapping functions provided with this IOMMU driver.
2604 * Must be called with the domain lock held.
2606 static dma_addr_t
__map_single(struct device
*dev
,
2607 struct dma_ops_domain
*dma_dom
,
2614 dma_addr_t offset
= paddr
& ~PAGE_MASK
;
2615 dma_addr_t address
, start
, ret
;
2617 unsigned long align_mask
= 0;
2620 pages
= iommu_num_pages(paddr
, size
, PAGE_SIZE
);
2623 INC_STATS_COUNTER(total_map_requests
);
2626 INC_STATS_COUNTER(cross_page
);
2629 align_mask
= (1UL << get_order(size
)) - 1;
2632 address
= dma_ops_alloc_addresses(dev
, dma_dom
, pages
, align_mask
,
2634 if (unlikely(address
== DMA_ERROR_CODE
)) {
2636 * setting next_address here will let the address
2637 * allocator only scan the new allocated range in the
2638 * first run. This is a small optimization.
2640 dma_dom
->next_address
= dma_dom
->aperture_size
;
2642 if (alloc_new_range(dma_dom
, false, GFP_ATOMIC
))
2646 * aperture was successfully enlarged by 128 MB, try
2653 for (i
= 0; i
< pages
; ++i
) {
2654 ret
= dma_ops_domain_map(dma_dom
, start
, paddr
, dir
);
2655 if (ret
== DMA_ERROR_CODE
)
2663 ADD_STATS_COUNTER(alloced_io_mem
, size
);
2665 if (unlikely(dma_dom
->need_flush
&& !amd_iommu_unmap_flush
)) {
2666 domain_flush_tlb(&dma_dom
->domain
);
2667 dma_dom
->need_flush
= false;
2668 } else if (unlikely(amd_iommu_np_cache
))
2669 domain_flush_pages(&dma_dom
->domain
, address
, size
);
2676 for (--i
; i
>= 0; --i
) {
2678 dma_ops_domain_unmap(dma_dom
, start
);
2681 dma_ops_free_addresses(dma_dom
, address
, pages
);
2683 return DMA_ERROR_CODE
;
2687 * Does the reverse of the __map_single function. Must be called with
2688 * the domain lock held too
2690 static void __unmap_single(struct dma_ops_domain
*dma_dom
,
2691 dma_addr_t dma_addr
,
2695 dma_addr_t flush_addr
;
2696 dma_addr_t i
, start
;
2699 if ((dma_addr
== DMA_ERROR_CODE
) ||
2700 (dma_addr
+ size
> dma_dom
->aperture_size
))
2703 flush_addr
= dma_addr
;
2704 pages
= iommu_num_pages(dma_addr
, size
, PAGE_SIZE
);
2705 dma_addr
&= PAGE_MASK
;
2708 for (i
= 0; i
< pages
; ++i
) {
2709 dma_ops_domain_unmap(dma_dom
, start
);
2713 SUB_STATS_COUNTER(alloced_io_mem
, size
);
2715 dma_ops_free_addresses(dma_dom
, dma_addr
, pages
);
2717 if (amd_iommu_unmap_flush
|| dma_dom
->need_flush
) {
2718 domain_flush_pages(&dma_dom
->domain
, flush_addr
, size
);
2719 dma_dom
->need_flush
= false;
2724 * The exported map_single function for dma_ops.
2726 static dma_addr_t
map_page(struct device
*dev
, struct page
*page
,
2727 unsigned long offset
, size_t size
,
2728 enum dma_data_direction dir
,
2729 struct dma_attrs
*attrs
)
2731 unsigned long flags
;
2732 struct protection_domain
*domain
;
2735 phys_addr_t paddr
= page_to_phys(page
) + offset
;
2737 INC_STATS_COUNTER(cnt_map_single
);
2739 domain
= get_domain(dev
);
2740 if (PTR_ERR(domain
) == -EINVAL
)
2741 return (dma_addr_t
)paddr
;
2742 else if (IS_ERR(domain
))
2743 return DMA_ERROR_CODE
;
2745 dma_mask
= *dev
->dma_mask
;
2747 spin_lock_irqsave(&domain
->lock
, flags
);
2749 addr
= __map_single(dev
, domain
->priv
, paddr
, size
, dir
, false,
2751 if (addr
== DMA_ERROR_CODE
)
2754 domain_flush_complete(domain
);
2757 spin_unlock_irqrestore(&domain
->lock
, flags
);
2763 * The exported unmap_single function for dma_ops.
2765 static void unmap_page(struct device
*dev
, dma_addr_t dma_addr
, size_t size
,
2766 enum dma_data_direction dir
, struct dma_attrs
*attrs
)
2768 unsigned long flags
;
2769 struct protection_domain
*domain
;
2771 INC_STATS_COUNTER(cnt_unmap_single
);
2773 domain
= get_domain(dev
);
2777 spin_lock_irqsave(&domain
->lock
, flags
);
2779 __unmap_single(domain
->priv
, dma_addr
, size
, dir
);
2781 domain_flush_complete(domain
);
2783 spin_unlock_irqrestore(&domain
->lock
, flags
);
2787 * This is a special map_sg function which is used if we should map a
2788 * device which is not handled by an AMD IOMMU in the system.
2790 static int map_sg_no_iommu(struct device
*dev
, struct scatterlist
*sglist
,
2791 int nelems
, int dir
)
2793 struct scatterlist
*s
;
2796 for_each_sg(sglist
, s
, nelems
, i
) {
2797 s
->dma_address
= (dma_addr_t
)sg_phys(s
);
2798 s
->dma_length
= s
->length
;
2805 * The exported map_sg function for dma_ops (handles scatter-gather
2808 static int map_sg(struct device
*dev
, struct scatterlist
*sglist
,
2809 int nelems
, enum dma_data_direction dir
,
2810 struct dma_attrs
*attrs
)
2812 unsigned long flags
;
2813 struct protection_domain
*domain
;
2815 struct scatterlist
*s
;
2817 int mapped_elems
= 0;
2820 INC_STATS_COUNTER(cnt_map_sg
);
2822 domain
= get_domain(dev
);
2823 if (PTR_ERR(domain
) == -EINVAL
)
2824 return map_sg_no_iommu(dev
, sglist
, nelems
, dir
);
2825 else if (IS_ERR(domain
))
2828 dma_mask
= *dev
->dma_mask
;
2830 spin_lock_irqsave(&domain
->lock
, flags
);
2832 for_each_sg(sglist
, s
, nelems
, i
) {
2835 s
->dma_address
= __map_single(dev
, domain
->priv
,
2836 paddr
, s
->length
, dir
, false,
2839 if (s
->dma_address
) {
2840 s
->dma_length
= s
->length
;
2846 domain_flush_complete(domain
);
2849 spin_unlock_irqrestore(&domain
->lock
, flags
);
2851 return mapped_elems
;
2853 for_each_sg(sglist
, s
, mapped_elems
, i
) {
2855 __unmap_single(domain
->priv
, s
->dma_address
,
2856 s
->dma_length
, dir
);
2857 s
->dma_address
= s
->dma_length
= 0;
2866 * The exported map_sg function for dma_ops (handles scatter-gather
2869 static void unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
2870 int nelems
, enum dma_data_direction dir
,
2871 struct dma_attrs
*attrs
)
2873 unsigned long flags
;
2874 struct protection_domain
*domain
;
2875 struct scatterlist
*s
;
2878 INC_STATS_COUNTER(cnt_unmap_sg
);
2880 domain
= get_domain(dev
);
2884 spin_lock_irqsave(&domain
->lock
, flags
);
2886 for_each_sg(sglist
, s
, nelems
, i
) {
2887 __unmap_single(domain
->priv
, s
->dma_address
,
2888 s
->dma_length
, dir
);
2889 s
->dma_address
= s
->dma_length
= 0;
2892 domain_flush_complete(domain
);
2894 spin_unlock_irqrestore(&domain
->lock
, flags
);
2898 * The exported alloc_coherent function for dma_ops.
2900 static void *alloc_coherent(struct device
*dev
, size_t size
,
2901 dma_addr_t
*dma_addr
, gfp_t flag
,
2902 struct dma_attrs
*attrs
)
2904 unsigned long flags
;
2906 struct protection_domain
*domain
;
2908 u64 dma_mask
= dev
->coherent_dma_mask
;
2910 INC_STATS_COUNTER(cnt_alloc_coherent
);
2912 domain
= get_domain(dev
);
2913 if (PTR_ERR(domain
) == -EINVAL
) {
2914 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2915 *dma_addr
= __pa(virt_addr
);
2917 } else if (IS_ERR(domain
))
2920 dma_mask
= dev
->coherent_dma_mask
;
2921 flag
&= ~(__GFP_DMA
| __GFP_HIGHMEM
| __GFP_DMA32
);
2924 virt_addr
= (void *)__get_free_pages(flag
, get_order(size
));
2928 paddr
= virt_to_phys(virt_addr
);
2931 dma_mask
= *dev
->dma_mask
;
2933 spin_lock_irqsave(&domain
->lock
, flags
);
2935 *dma_addr
= __map_single(dev
, domain
->priv
, paddr
,
2936 size
, DMA_BIDIRECTIONAL
, true, dma_mask
);
2938 if (*dma_addr
== DMA_ERROR_CODE
) {
2939 spin_unlock_irqrestore(&domain
->lock
, flags
);
2943 domain_flush_complete(domain
);
2945 spin_unlock_irqrestore(&domain
->lock
, flags
);
2951 free_pages((unsigned long)virt_addr
, get_order(size
));
2957 * The exported free_coherent function for dma_ops.
2959 static void free_coherent(struct device
*dev
, size_t size
,
2960 void *virt_addr
, dma_addr_t dma_addr
,
2961 struct dma_attrs
*attrs
)
2963 unsigned long flags
;
2964 struct protection_domain
*domain
;
2966 INC_STATS_COUNTER(cnt_free_coherent
);
2968 domain
= get_domain(dev
);
2972 spin_lock_irqsave(&domain
->lock
, flags
);
2974 __unmap_single(domain
->priv
, dma_addr
, size
, DMA_BIDIRECTIONAL
);
2976 domain_flush_complete(domain
);
2978 spin_unlock_irqrestore(&domain
->lock
, flags
);
2981 free_pages((unsigned long)virt_addr
, get_order(size
));
2985 * This function is called by the DMA layer to find out if we can handle a
2986 * particular device. It is part of the dma_ops.
2988 static int amd_iommu_dma_supported(struct device
*dev
, u64 mask
)
2990 return check_device(dev
);
2994 * The function for pre-allocating protection domains.
2996 * If the driver core informs the DMA layer if a driver grabs a device
2997 * we don't need to preallocate the protection domains anymore.
2998 * For now we have to.
3000 static void __init
prealloc_protection_domains(void)
3002 struct iommu_dev_data
*dev_data
;
3003 struct dma_ops_domain
*dma_dom
;
3004 struct pci_dev
*dev
= NULL
;
3007 for_each_pci_dev(dev
) {
3009 /* Do we handle this device? */
3010 if (!check_device(&dev
->dev
))
3013 dev_data
= get_dev_data(&dev
->dev
);
3014 if (!amd_iommu_force_isolation
&& dev_data
->iommu_v2
) {
3015 /* Make sure passthrough domain is allocated */
3016 alloc_passthrough_domain();
3017 dev_data
->passthrough
= true;
3018 attach_device(&dev
->dev
, pt_domain
);
3019 pr_info("AMD-Vi: Using passthrough domain for device %s\n",
3020 dev_name(&dev
->dev
));
3023 /* Is there already any domain for it? */
3024 if (domain_for_device(&dev
->dev
))
3027 devid
= get_device_id(&dev
->dev
);
3029 dma_dom
= dma_ops_domain_alloc();
3032 init_unity_mappings_for_device(dma_dom
, devid
);
3033 dma_dom
->target_dev
= devid
;
3035 attach_device(&dev
->dev
, &dma_dom
->domain
);
3037 list_add_tail(&dma_dom
->list
, &iommu_pd_list
);
3041 static struct dma_map_ops amd_iommu_dma_ops
= {
3042 .alloc
= alloc_coherent
,
3043 .free
= free_coherent
,
3044 .map_page
= map_page
,
3045 .unmap_page
= unmap_page
,
3047 .unmap_sg
= unmap_sg
,
3048 .dma_supported
= amd_iommu_dma_supported
,
3051 static unsigned device_dma_ops_init(void)
3053 struct iommu_dev_data
*dev_data
;
3054 struct pci_dev
*pdev
= NULL
;
3055 unsigned unhandled
= 0;
3057 for_each_pci_dev(pdev
) {
3058 if (!check_device(&pdev
->dev
)) {
3060 iommu_ignore_device(&pdev
->dev
);
3066 dev_data
= get_dev_data(&pdev
->dev
);
3068 if (!dev_data
->passthrough
)
3069 pdev
->dev
.archdata
.dma_ops
= &amd_iommu_dma_ops
;
3071 pdev
->dev
.archdata
.dma_ops
= &nommu_dma_ops
;
3078 * The function which clues the AMD IOMMU driver into dma_ops.
3081 void __init
amd_iommu_init_api(void)
3083 bus_set_iommu(&pci_bus_type
, &amd_iommu_ops
);
3086 int __init
amd_iommu_init_dma_ops(void)
3088 struct amd_iommu
*iommu
;
3092 * first allocate a default protection domain for every IOMMU we
3093 * found in the system. Devices not assigned to any other
3094 * protection domain will be assigned to the default one.
3096 for_each_iommu(iommu
) {
3097 iommu
->default_dom
= dma_ops_domain_alloc();
3098 if (iommu
->default_dom
== NULL
)
3100 iommu
->default_dom
->domain
.flags
|= PD_DEFAULT_MASK
;
3101 ret
= iommu_init_unity_mappings(iommu
);
3107 * Pre-allocate the protection domains for each device.
3109 prealloc_protection_domains();
3114 /* Make the driver finally visible to the drivers */
3115 unhandled
= device_dma_ops_init();
3116 if (unhandled
&& max_pfn
> MAX_DMA32_PFN
) {
3117 /* There are unhandled devices - initialize swiotlb for them */
3121 amd_iommu_stats_init();
3123 if (amd_iommu_unmap_flush
)
3124 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
3126 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
3132 for_each_iommu(iommu
) {
3133 if (iommu
->default_dom
)
3134 dma_ops_domain_free(iommu
->default_dom
);
3140 /*****************************************************************************
3142 * The following functions belong to the exported interface of AMD IOMMU
3144 * This interface allows access to lower level functions of the IOMMU
3145 * like protection domain handling and assignement of devices to domains
3146 * which is not possible with the dma_ops interface.
3148 *****************************************************************************/
3150 static void cleanup_domain(struct protection_domain
*domain
)
3152 struct iommu_dev_data
*dev_data
, *next
;
3153 unsigned long flags
;
3155 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
3157 list_for_each_entry_safe(dev_data
, next
, &domain
->dev_list
, list
) {
3158 __detach_device(dev_data
);
3159 atomic_set(&dev_data
->bind
, 0);
3162 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
3165 static void protection_domain_free(struct protection_domain
*domain
)
3170 del_domain_from_list(domain
);
3173 domain_id_free(domain
->id
);
3178 static struct protection_domain
*protection_domain_alloc(void)
3180 struct protection_domain
*domain
;
3182 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
3186 spin_lock_init(&domain
->lock
);
3187 mutex_init(&domain
->api_lock
);
3188 domain
->id
= domain_id_alloc();
3191 INIT_LIST_HEAD(&domain
->dev_list
);
3193 add_domain_to_list(domain
);
3203 static int __init
alloc_passthrough_domain(void)
3205 if (pt_domain
!= NULL
)
3208 /* allocate passthrough domain */
3209 pt_domain
= protection_domain_alloc();
3213 pt_domain
->mode
= PAGE_MODE_NONE
;
3217 static int amd_iommu_domain_init(struct iommu_domain
*dom
)
3219 struct protection_domain
*domain
;
3221 domain
= protection_domain_alloc();
3225 domain
->mode
= PAGE_MODE_3_LEVEL
;
3226 domain
->pt_root
= (void *)get_zeroed_page(GFP_KERNEL
);
3227 if (!domain
->pt_root
)
3230 domain
->iommu_domain
= dom
;
3234 dom
->geometry
.aperture_start
= 0;
3235 dom
->geometry
.aperture_end
= ~0ULL;
3236 dom
->geometry
.force_aperture
= true;
3241 protection_domain_free(domain
);
3246 static void amd_iommu_domain_destroy(struct iommu_domain
*dom
)
3248 struct protection_domain
*domain
= dom
->priv
;
3253 if (domain
->dev_cnt
> 0)
3254 cleanup_domain(domain
);
3256 BUG_ON(domain
->dev_cnt
!= 0);
3258 if (domain
->mode
!= PAGE_MODE_NONE
)
3259 free_pagetable(domain
);
3261 if (domain
->flags
& PD_IOMMUV2_MASK
)
3262 free_gcr3_table(domain
);
3264 protection_domain_free(domain
);
3269 static void amd_iommu_detach_device(struct iommu_domain
*dom
,
3272 struct iommu_dev_data
*dev_data
= dev
->archdata
.iommu
;
3273 struct amd_iommu
*iommu
;
3276 if (!check_device(dev
))
3279 devid
= get_device_id(dev
);
3281 if (dev_data
->domain
!= NULL
)
3284 iommu
= amd_iommu_rlookup_table
[devid
];
3288 iommu_completion_wait(iommu
);
3291 static int amd_iommu_attach_device(struct iommu_domain
*dom
,
3294 struct protection_domain
*domain
= dom
->priv
;
3295 struct iommu_dev_data
*dev_data
;
3296 struct amd_iommu
*iommu
;
3299 if (!check_device(dev
))
3302 dev_data
= dev
->archdata
.iommu
;
3304 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3308 if (dev_data
->domain
)
3311 ret
= attach_device(dev
, domain
);
3313 iommu_completion_wait(iommu
);
3318 static int amd_iommu_map(struct iommu_domain
*dom
, unsigned long iova
,
3319 phys_addr_t paddr
, size_t page_size
, int iommu_prot
)
3321 struct protection_domain
*domain
= dom
->priv
;
3325 if (domain
->mode
== PAGE_MODE_NONE
)
3328 if (iommu_prot
& IOMMU_READ
)
3329 prot
|= IOMMU_PROT_IR
;
3330 if (iommu_prot
& IOMMU_WRITE
)
3331 prot
|= IOMMU_PROT_IW
;
3333 mutex_lock(&domain
->api_lock
);
3334 ret
= iommu_map_page(domain
, iova
, paddr
, prot
, page_size
);
3335 mutex_unlock(&domain
->api_lock
);
3340 static size_t amd_iommu_unmap(struct iommu_domain
*dom
, unsigned long iova
,
3343 struct protection_domain
*domain
= dom
->priv
;
3346 if (domain
->mode
== PAGE_MODE_NONE
)
3349 mutex_lock(&domain
->api_lock
);
3350 unmap_size
= iommu_unmap_page(domain
, iova
, page_size
);
3351 mutex_unlock(&domain
->api_lock
);
3353 domain_flush_tlb_pde(domain
);
3358 static phys_addr_t
amd_iommu_iova_to_phys(struct iommu_domain
*dom
,
3361 struct protection_domain
*domain
= dom
->priv
;
3362 unsigned long offset_mask
;
3366 if (domain
->mode
== PAGE_MODE_NONE
)
3369 pte
= fetch_pte(domain
, iova
);
3371 if (!pte
|| !IOMMU_PTE_PRESENT(*pte
))
3374 if (PM_PTE_LEVEL(*pte
) == 0)
3375 offset_mask
= PAGE_SIZE
- 1;
3377 offset_mask
= PTE_PAGE_SIZE(*pte
) - 1;
3379 __pte
= *pte
& PM_ADDR_MASK
;
3380 paddr
= (__pte
& ~offset_mask
) | (iova
& offset_mask
);
3385 static int amd_iommu_domain_has_cap(struct iommu_domain
*domain
,
3389 case IOMMU_CAP_CACHE_COHERENCY
:
3391 case IOMMU_CAP_INTR_REMAP
:
3392 return irq_remapping_enabled
;
3398 static struct iommu_ops amd_iommu_ops
= {
3399 .domain_init
= amd_iommu_domain_init
,
3400 .domain_destroy
= amd_iommu_domain_destroy
,
3401 .attach_dev
= amd_iommu_attach_device
,
3402 .detach_dev
= amd_iommu_detach_device
,
3403 .map
= amd_iommu_map
,
3404 .unmap
= amd_iommu_unmap
,
3405 .iova_to_phys
= amd_iommu_iova_to_phys
,
3406 .domain_has_cap
= amd_iommu_domain_has_cap
,
3407 .pgsize_bitmap
= AMD_IOMMU_PGSIZES
,
3410 /*****************************************************************************
3412 * The next functions do a basic initialization of IOMMU for pass through
3415 * In passthrough mode the IOMMU is initialized and enabled but not used for
3416 * DMA-API translation.
3418 *****************************************************************************/
3420 int __init
amd_iommu_init_passthrough(void)
3422 struct iommu_dev_data
*dev_data
;
3423 struct pci_dev
*dev
= NULL
;
3424 struct amd_iommu
*iommu
;
3428 ret
= alloc_passthrough_domain();
3432 for_each_pci_dev(dev
) {
3433 if (!check_device(&dev
->dev
))
3436 dev_data
= get_dev_data(&dev
->dev
);
3437 dev_data
->passthrough
= true;
3439 devid
= get_device_id(&dev
->dev
);
3441 iommu
= amd_iommu_rlookup_table
[devid
];
3445 attach_device(&dev
->dev
, pt_domain
);
3448 amd_iommu_stats_init();
3450 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3455 /* IOMMUv2 specific functions */
3456 int amd_iommu_register_ppr_notifier(struct notifier_block
*nb
)
3458 return atomic_notifier_chain_register(&ppr_notifier
, nb
);
3460 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier
);
3462 int amd_iommu_unregister_ppr_notifier(struct notifier_block
*nb
)
3464 return atomic_notifier_chain_unregister(&ppr_notifier
, nb
);
3466 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier
);
3468 void amd_iommu_domain_direct_map(struct iommu_domain
*dom
)
3470 struct protection_domain
*domain
= dom
->priv
;
3471 unsigned long flags
;
3473 spin_lock_irqsave(&domain
->lock
, flags
);
3475 /* Update data structure */
3476 domain
->mode
= PAGE_MODE_NONE
;
3477 domain
->updated
= true;
3479 /* Make changes visible to IOMMUs */
3480 update_domain(domain
);
3482 /* Page-table is not visible to IOMMU anymore, so free it */
3483 free_pagetable(domain
);
3485 spin_unlock_irqrestore(&domain
->lock
, flags
);
3487 EXPORT_SYMBOL(amd_iommu_domain_direct_map
);
3489 int amd_iommu_domain_enable_v2(struct iommu_domain
*dom
, int pasids
)
3491 struct protection_domain
*domain
= dom
->priv
;
3492 unsigned long flags
;
3495 if (pasids
<= 0 || pasids
> (PASID_MASK
+ 1))
3498 /* Number of GCR3 table levels required */
3499 for (levels
= 0; (pasids
- 1) & ~0x1ff; pasids
>>= 9)
3502 if (levels
> amd_iommu_max_glx_val
)
3505 spin_lock_irqsave(&domain
->lock
, flags
);
3508 * Save us all sanity checks whether devices already in the
3509 * domain support IOMMUv2. Just force that the domain has no
3510 * devices attached when it is switched into IOMMUv2 mode.
3513 if (domain
->dev_cnt
> 0 || domain
->flags
& PD_IOMMUV2_MASK
)
3517 domain
->gcr3_tbl
= (void *)get_zeroed_page(GFP_ATOMIC
);
3518 if (domain
->gcr3_tbl
== NULL
)
3521 domain
->glx
= levels
;
3522 domain
->flags
|= PD_IOMMUV2_MASK
;
3523 domain
->updated
= true;
3525 update_domain(domain
);
3530 spin_unlock_irqrestore(&domain
->lock
, flags
);
3534 EXPORT_SYMBOL(amd_iommu_domain_enable_v2
);
3536 static int __flush_pasid(struct protection_domain
*domain
, int pasid
,
3537 u64 address
, bool size
)
3539 struct iommu_dev_data
*dev_data
;
3540 struct iommu_cmd cmd
;
3543 if (!(domain
->flags
& PD_IOMMUV2_MASK
))
3546 build_inv_iommu_pasid(&cmd
, domain
->id
, pasid
, address
, size
);
3549 * IOMMU TLB needs to be flushed before Device TLB to
3550 * prevent device TLB refill from IOMMU TLB
3552 for (i
= 0; i
< amd_iommus_present
; ++i
) {
3553 if (domain
->dev_iommu
[i
] == 0)
3556 ret
= iommu_queue_command(amd_iommus
[i
], &cmd
);
3561 /* Wait until IOMMU TLB flushes are complete */
3562 domain_flush_complete(domain
);
3564 /* Now flush device TLBs */
3565 list_for_each_entry(dev_data
, &domain
->dev_list
, list
) {
3566 struct amd_iommu
*iommu
;
3569 BUG_ON(!dev_data
->ats
.enabled
);
3571 qdep
= dev_data
->ats
.qdep
;
3572 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3574 build_inv_iotlb_pasid(&cmd
, dev_data
->devid
, pasid
,
3575 qdep
, address
, size
);
3577 ret
= iommu_queue_command(iommu
, &cmd
);
3582 /* Wait until all device TLBs are flushed */
3583 domain_flush_complete(domain
);
3592 static int __amd_iommu_flush_page(struct protection_domain
*domain
, int pasid
,
3595 INC_STATS_COUNTER(invalidate_iotlb
);
3597 return __flush_pasid(domain
, pasid
, address
, false);
3600 int amd_iommu_flush_page(struct iommu_domain
*dom
, int pasid
,
3603 struct protection_domain
*domain
= dom
->priv
;
3604 unsigned long flags
;
3607 spin_lock_irqsave(&domain
->lock
, flags
);
3608 ret
= __amd_iommu_flush_page(domain
, pasid
, address
);
3609 spin_unlock_irqrestore(&domain
->lock
, flags
);
3613 EXPORT_SYMBOL(amd_iommu_flush_page
);
3615 static int __amd_iommu_flush_tlb(struct protection_domain
*domain
, int pasid
)
3617 INC_STATS_COUNTER(invalidate_iotlb_all
);
3619 return __flush_pasid(domain
, pasid
, CMD_INV_IOMMU_ALL_PAGES_ADDRESS
,
3623 int amd_iommu_flush_tlb(struct iommu_domain
*dom
, int pasid
)
3625 struct protection_domain
*domain
= dom
->priv
;
3626 unsigned long flags
;
3629 spin_lock_irqsave(&domain
->lock
, flags
);
3630 ret
= __amd_iommu_flush_tlb(domain
, pasid
);
3631 spin_unlock_irqrestore(&domain
->lock
, flags
);
3635 EXPORT_SYMBOL(amd_iommu_flush_tlb
);
3637 static u64
*__get_gcr3_pte(u64
*root
, int level
, int pasid
, bool alloc
)
3644 index
= (pasid
>> (9 * level
)) & 0x1ff;
3650 if (!(*pte
& GCR3_VALID
)) {
3654 root
= (void *)get_zeroed_page(GFP_ATOMIC
);
3658 *pte
= __pa(root
) | GCR3_VALID
;
3661 root
= __va(*pte
& PAGE_MASK
);
3669 static int __set_gcr3(struct protection_domain
*domain
, int pasid
,
3674 if (domain
->mode
!= PAGE_MODE_NONE
)
3677 pte
= __get_gcr3_pte(domain
->gcr3_tbl
, domain
->glx
, pasid
, true);
3681 *pte
= (cr3
& PAGE_MASK
) | GCR3_VALID
;
3683 return __amd_iommu_flush_tlb(domain
, pasid
);
3686 static int __clear_gcr3(struct protection_domain
*domain
, int pasid
)
3690 if (domain
->mode
!= PAGE_MODE_NONE
)
3693 pte
= __get_gcr3_pte(domain
->gcr3_tbl
, domain
->glx
, pasid
, false);
3699 return __amd_iommu_flush_tlb(domain
, pasid
);
3702 int amd_iommu_domain_set_gcr3(struct iommu_domain
*dom
, int pasid
,
3705 struct protection_domain
*domain
= dom
->priv
;
3706 unsigned long flags
;
3709 spin_lock_irqsave(&domain
->lock
, flags
);
3710 ret
= __set_gcr3(domain
, pasid
, cr3
);
3711 spin_unlock_irqrestore(&domain
->lock
, flags
);
3715 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3
);
3717 int amd_iommu_domain_clear_gcr3(struct iommu_domain
*dom
, int pasid
)
3719 struct protection_domain
*domain
= dom
->priv
;
3720 unsigned long flags
;
3723 spin_lock_irqsave(&domain
->lock
, flags
);
3724 ret
= __clear_gcr3(domain
, pasid
);
3725 spin_unlock_irqrestore(&domain
->lock
, flags
);
3729 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3
);
3731 int amd_iommu_complete_ppr(struct pci_dev
*pdev
, int pasid
,
3732 int status
, int tag
)
3734 struct iommu_dev_data
*dev_data
;
3735 struct amd_iommu
*iommu
;
3736 struct iommu_cmd cmd
;
3738 INC_STATS_COUNTER(complete_ppr
);
3740 dev_data
= get_dev_data(&pdev
->dev
);
3741 iommu
= amd_iommu_rlookup_table
[dev_data
->devid
];
3743 build_complete_ppr(&cmd
, dev_data
->devid
, pasid
, status
,
3744 tag
, dev_data
->pri_tlp
);
3746 return iommu_queue_command(iommu
, &cmd
);
3748 EXPORT_SYMBOL(amd_iommu_complete_ppr
);
3750 struct iommu_domain
*amd_iommu_get_v2_domain(struct pci_dev
*pdev
)
3752 struct protection_domain
*domain
;
3754 domain
= get_domain(&pdev
->dev
);
3758 /* Only return IOMMUv2 domains */
3759 if (!(domain
->flags
& PD_IOMMUV2_MASK
))
3762 return domain
->iommu_domain
;
3764 EXPORT_SYMBOL(amd_iommu_get_v2_domain
);
3766 void amd_iommu_enable_device_erratum(struct pci_dev
*pdev
, u32 erratum
)
3768 struct iommu_dev_data
*dev_data
;
3770 if (!amd_iommu_v2_supported())
3773 dev_data
= get_dev_data(&pdev
->dev
);
3774 dev_data
->errata
|= (1 << erratum
);
3776 EXPORT_SYMBOL(amd_iommu_enable_device_erratum
);
3778 int amd_iommu_device_info(struct pci_dev
*pdev
,
3779 struct amd_iommu_device_info
*info
)
3784 if (pdev
== NULL
|| info
== NULL
)
3787 if (!amd_iommu_v2_supported())
3790 memset(info
, 0, sizeof(*info
));
3792 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_ATS
);
3794 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_ATS_SUP
;
3796 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PRI
);
3798 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PRI_SUP
;
3800 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_PASID
);
3804 max_pasids
= 1 << (9 * (amd_iommu_max_glx_val
+ 1));
3805 max_pasids
= min(max_pasids
, (1 << 20));
3807 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PASID_SUP
;
3808 info
->max_pasids
= min(pci_max_pasids(pdev
), max_pasids
);
3810 features
= pci_pasid_features(pdev
);
3811 if (features
& PCI_PASID_CAP_EXEC
)
3812 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP
;
3813 if (features
& PCI_PASID_CAP_PRIV
)
3814 info
->flags
|= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP
;
3819 EXPORT_SYMBOL(amd_iommu_device_info
);
3821 #ifdef CONFIG_IRQ_REMAP
3823 /*****************************************************************************
3825 * Interrupt Remapping Implementation
3827 *****************************************************************************/
3844 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3845 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3846 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3847 #define DTE_IRQ_REMAP_ENABLE 1ULL
3849 static void set_dte_irq_entry(u16 devid
, struct irq_remap_table
*table
)
3853 dte
= amd_iommu_dev_table
[devid
].data
[2];
3854 dte
&= ~DTE_IRQ_PHYS_ADDR_MASK
;
3855 dte
|= virt_to_phys(table
->table
);
3856 dte
|= DTE_IRQ_REMAP_INTCTL
;
3857 dte
|= DTE_IRQ_TABLE_LEN
;
3858 dte
|= DTE_IRQ_REMAP_ENABLE
;
3860 amd_iommu_dev_table
[devid
].data
[2] = dte
;
3863 #define IRTE_ALLOCATED (~1U)
3865 static struct irq_remap_table
*get_irq_table(u16 devid
, bool ioapic
)
3867 struct irq_remap_table
*table
= NULL
;
3868 struct amd_iommu
*iommu
;
3869 unsigned long flags
;
3872 write_lock_irqsave(&amd_iommu_devtable_lock
, flags
);
3874 iommu
= amd_iommu_rlookup_table
[devid
];
3878 table
= irq_lookup_table
[devid
];
3882 alias
= amd_iommu_alias_table
[devid
];
3883 table
= irq_lookup_table
[alias
];
3885 irq_lookup_table
[devid
] = table
;
3886 set_dte_irq_entry(devid
, table
);
3887 iommu_flush_dte(iommu
, devid
);
3891 /* Nothing there yet, allocate new irq remapping table */
3892 table
= kzalloc(sizeof(*table
), GFP_ATOMIC
);
3897 /* Keep the first 32 indexes free for IOAPIC interrupts */
3898 table
->min_index
= 32;
3900 table
->table
= kmem_cache_alloc(amd_iommu_irq_cache
, GFP_ATOMIC
);
3901 if (!table
->table
) {
3907 memset(table
->table
, 0, MAX_IRQS_PER_TABLE
* sizeof(u32
));
3912 for (i
= 0; i
< 32; ++i
)
3913 table
->table
[i
] = IRTE_ALLOCATED
;
3916 irq_lookup_table
[devid
] = table
;
3917 set_dte_irq_entry(devid
, table
);
3918 iommu_flush_dte(iommu
, devid
);
3919 if (devid
!= alias
) {
3920 irq_lookup_table
[alias
] = table
;
3921 set_dte_irq_entry(devid
, table
);
3922 iommu_flush_dte(iommu
, alias
);
3926 iommu_completion_wait(iommu
);
3929 write_unlock_irqrestore(&amd_iommu_devtable_lock
, flags
);
3934 static int alloc_irq_index(struct irq_cfg
*cfg
, u16 devid
, int count
)
3936 struct irq_remap_table
*table
;
3937 unsigned long flags
;
3940 table
= get_irq_table(devid
, false);
3944 spin_lock_irqsave(&table
->lock
, flags
);
3946 /* Scan table for free entries */
3947 for (c
= 0, index
= table
->min_index
;
3948 index
< MAX_IRQS_PER_TABLE
;
3950 if (table
->table
[index
] == 0)
3956 struct irq_2_iommu
*irte_info
;
3959 table
->table
[index
- c
+ 1] = IRTE_ALLOCATED
;
3963 irte_info
= &cfg
->irq_2_iommu
;
3964 irte_info
->sub_handle
= devid
;
3965 irte_info
->irte_index
= index
;
3966 irte_info
->iommu
= (void *)cfg
;
3975 spin_unlock_irqrestore(&table
->lock
, flags
);
3980 static int get_irte(u16 devid
, int index
, union irte
*irte
)
3982 struct irq_remap_table
*table
;
3983 unsigned long flags
;
3985 table
= get_irq_table(devid
, false);
3989 spin_lock_irqsave(&table
->lock
, flags
);
3990 irte
->val
= table
->table
[index
];
3991 spin_unlock_irqrestore(&table
->lock
, flags
);
3996 static int modify_irte(u16 devid
, int index
, union irte irte
)
3998 struct irq_remap_table
*table
;
3999 struct amd_iommu
*iommu
;
4000 unsigned long flags
;
4002 iommu
= amd_iommu_rlookup_table
[devid
];
4006 table
= get_irq_table(devid
, false);
4010 spin_lock_irqsave(&table
->lock
, flags
);
4011 table
->table
[index
] = irte
.val
;
4012 spin_unlock_irqrestore(&table
->lock
, flags
);
4014 iommu_flush_irt(iommu
, devid
);
4015 iommu_completion_wait(iommu
);
4020 static void free_irte(u16 devid
, int index
)
4022 struct irq_remap_table
*table
;
4023 struct amd_iommu
*iommu
;
4024 unsigned long flags
;
4026 iommu
= amd_iommu_rlookup_table
[devid
];
4030 table
= get_irq_table(devid
, false);
4034 spin_lock_irqsave(&table
->lock
, flags
);
4035 table
->table
[index
] = 0;
4036 spin_unlock_irqrestore(&table
->lock
, flags
);
4038 iommu_flush_irt(iommu
, devid
);
4039 iommu_completion_wait(iommu
);
4042 static int setup_ioapic_entry(int irq
, struct IO_APIC_route_entry
*entry
,
4043 unsigned int destination
, int vector
,
4044 struct io_apic_irq_attr
*attr
)
4046 struct irq_remap_table
*table
;
4047 struct irq_2_iommu
*irte_info
;
4048 struct irq_cfg
*cfg
;
4055 cfg
= irq_get_chip_data(irq
);
4059 irte_info
= &cfg
->irq_2_iommu
;
4060 ioapic_id
= mpc_ioapic_id(attr
->ioapic
);
4061 devid
= get_ioapic_devid(ioapic_id
);
4066 table
= get_irq_table(devid
, true);
4070 index
= attr
->ioapic_pin
;
4072 /* Setup IRQ remapping info */
4073 irte_info
->sub_handle
= devid
;
4074 irte_info
->irte_index
= index
;
4075 irte_info
->iommu
= (void *)cfg
;
4077 /* Setup IRTE for IOMMU */
4079 irte
.fields
.vector
= vector
;
4080 irte
.fields
.int_type
= apic
->irq_delivery_mode
;
4081 irte
.fields
.destination
= destination
;
4082 irte
.fields
.dm
= apic
->irq_dest_mode
;
4083 irte
.fields
.valid
= 1;
4085 ret
= modify_irte(devid
, index
, irte
);
4089 /* Setup IOAPIC entry */
4090 memset(entry
, 0, sizeof(*entry
));
4092 entry
->vector
= index
;
4094 entry
->trigger
= attr
->trigger
;
4095 entry
->polarity
= attr
->polarity
;
4098 * Mask level triggered irqs.
4106 static int set_affinity(struct irq_data
*data
, const struct cpumask
*mask
,
4109 struct irq_2_iommu
*irte_info
;
4110 unsigned int dest
, irq
;
4111 struct irq_cfg
*cfg
;
4115 if (!config_enabled(CONFIG_SMP
))
4118 cfg
= data
->chip_data
;
4120 irte_info
= &cfg
->irq_2_iommu
;
4122 if (!cpumask_intersects(mask
, cpu_online_mask
))
4125 if (get_irte(irte_info
->sub_handle
, irte_info
->irte_index
, &irte
))
4128 if (assign_irq_vector(irq
, cfg
, mask
))
4131 err
= apic
->cpu_mask_to_apicid_and(cfg
->domain
, mask
, &dest
);
4133 if (assign_irq_vector(irq
, cfg
, data
->affinity
))
4134 pr_err("AMD-Vi: Failed to recover vector for irq %d\n", irq
);
4138 irte
.fields
.vector
= cfg
->vector
;
4139 irte
.fields
.destination
= dest
;
4141 modify_irte(irte_info
->sub_handle
, irte_info
->irte_index
, irte
);
4143 if (cfg
->move_in_progress
)
4144 send_cleanup_vector(cfg
);
4146 cpumask_copy(data
->affinity
, mask
);
4151 static int free_irq(int irq
)
4153 struct irq_2_iommu
*irte_info
;
4154 struct irq_cfg
*cfg
;
4156 cfg
= irq_get_chip_data(irq
);
4160 irte_info
= &cfg
->irq_2_iommu
;
4162 free_irte(irte_info
->sub_handle
, irte_info
->irte_index
);
4167 static void compose_msi_msg(struct pci_dev
*pdev
,
4168 unsigned int irq
, unsigned int dest
,
4169 struct msi_msg
*msg
, u8 hpet_id
)
4171 struct irq_2_iommu
*irte_info
;
4172 struct irq_cfg
*cfg
;
4175 cfg
= irq_get_chip_data(irq
);
4179 irte_info
= &cfg
->irq_2_iommu
;
4182 irte
.fields
.vector
= cfg
->vector
;
4183 irte
.fields
.int_type
= apic
->irq_delivery_mode
;
4184 irte
.fields
.destination
= dest
;
4185 irte
.fields
.dm
= apic
->irq_dest_mode
;
4186 irte
.fields
.valid
= 1;
4188 modify_irte(irte_info
->sub_handle
, irte_info
->irte_index
, irte
);
4190 msg
->address_hi
= MSI_ADDR_BASE_HI
;
4191 msg
->address_lo
= MSI_ADDR_BASE_LO
;
4192 msg
->data
= irte_info
->irte_index
;
4195 static int msi_alloc_irq(struct pci_dev
*pdev
, int irq
, int nvec
)
4197 struct irq_cfg
*cfg
;
4204 cfg
= irq_get_chip_data(irq
);
4208 devid
= get_device_id(&pdev
->dev
);
4209 index
= alloc_irq_index(cfg
, devid
, nvec
);
4211 return index
< 0 ? MAX_IRQS_PER_TABLE
: index
;
4214 static int msi_setup_irq(struct pci_dev
*pdev
, unsigned int irq
,
4215 int index
, int offset
)
4217 struct irq_2_iommu
*irte_info
;
4218 struct irq_cfg
*cfg
;
4224 cfg
= irq_get_chip_data(irq
);
4228 if (index
>= MAX_IRQS_PER_TABLE
)
4231 devid
= get_device_id(&pdev
->dev
);
4232 irte_info
= &cfg
->irq_2_iommu
;
4234 irte_info
->sub_handle
= devid
;
4235 irte_info
->irte_index
= index
+ offset
;
4236 irte_info
->iommu
= (void *)cfg
;
4241 static int setup_hpet_msi(unsigned int irq
, unsigned int id
)
4243 struct irq_2_iommu
*irte_info
;
4244 struct irq_cfg
*cfg
;
4247 cfg
= irq_get_chip_data(irq
);
4251 irte_info
= &cfg
->irq_2_iommu
;
4252 devid
= get_hpet_devid(id
);
4256 index
= alloc_irq_index(cfg
, devid
, 1);
4260 irte_info
->sub_handle
= devid
;
4261 irte_info
->irte_index
= index
;
4262 irte_info
->iommu
= (void *)cfg
;
4267 struct irq_remap_ops amd_iommu_irq_ops
= {
4268 .supported
= amd_iommu_supported
,
4269 .prepare
= amd_iommu_prepare
,
4270 .enable
= amd_iommu_enable
,
4271 .disable
= amd_iommu_disable
,
4272 .reenable
= amd_iommu_reenable
,
4273 .enable_faulting
= amd_iommu_enable_faulting
,
4274 .setup_ioapic_entry
= setup_ioapic_entry
,
4275 .set_affinity
= set_affinity
,
4276 .free_irq
= free_irq
,
4277 .compose_msi_msg
= compose_msi_msg
,
4278 .msi_alloc_irq
= msi_alloc_irq
,
4279 .msi_setup_irq
= msi_setup_irq
,
4280 .setup_hpet_msi
= setup_hpet_msi
,